feat(ui): handle PresenceList in lobby — show online users

The lobby now populates from PresenceList signal events:
- Relay broadcasts user list on register/deregister
- JS receives "presence_list" signal-event
- Updates lobbyUsers map (excluding self)
- Renders user rows with identicon, name, fingerprint

Users appear in the lobby as soon as they register their
signal channel — no need to join voice first.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Cargo.toml

@@ -10,6 +10,8 @@ members = [
     "crates/wzp-client",
     "crates/wzp-web",
     "crates/wzp-android",
+    "crates/wzp-native",
+    "desktop/src-tauri",
 ]
 
 [workspace.package]
@@ -30,12 +32,20 @@ serde = { version = "1", features = ["derive"] }
 
 # Transport
 quinn = "0.11"
+socket2 = "0.5"
 
 # FEC
 raptorq = "2"
 
 # Codec
-audiopus = "0.3.0-rc.0"
+# opusic-c: high-level safe bindings over libopus 1.5.2 (encoder side).
+# opusic-sys: raw FFI for the decoder side — we build our own DecoderHandle
+# because opusic-c::Decoder.inner is pub(crate) and cannot be reached for the
+# Phase 3 DRED reconstruction path. See docs/PRD-dred-integration.md.
+# Pinned exactly (no caret) for reproducible libopus 1.5.2 across the fleet.
+opusic-c = { version = "=1.5.5", default-features = false, features = ["bundled", "dred"] }
+opusic-sys = { version = "=0.6.0", default-features = false, features = ["bundled"] }
+bytemuck = "1"
 codec2 = "0.3"
 
 # Crypto
@@ -53,3 +63,29 @@ wzp-fec = { path = "crates/wzp-fec" }
 wzp-crypto = { path = "crates/wzp-crypto" }
 wzp-transport = { path = "crates/wzp-transport" }
 wzp-client = { path = "crates/wzp-client" }
+
+# Fast dev profile: optimized but with debug info and incremental compilation.
+# Use with: cargo run --profile dev-fast
+[profile.dev-fast]
+inherits = "dev"
+opt-level = 2
+
+# Optimize heavy compute deps even in debug builds —
+# real-time audio needs < 20ms per frame, impossible unoptimized.
+[profile.dev.package.nnnoiseless]
+opt-level = 3
+[profile.dev.package.opusic-sys]
+opt-level = 3
+[profile.dev.package.raptorq]
+opt-level = 3
+[profile.dev.package.wzp-codec]
+opt-level = 3
+[profile.dev.package.wzp-fec]
+opt-level = 3
+
+# Phase 0 (opus-DRED): removed the [patch.crates-io] audiopus_sys = { path =
+# "vendor/audiopus_sys" } block. That patch existed to fix a Windows clang-cl
+# SIMD compile bug in libopus 1.3.1. With the swap to opusic-sys (libopus
+# 1.5.2), the upstream SIMD gating was fixed and the vendor patch is
+# obsolete. The vendor/audiopus_sys directory itself should be deleted as
+# part of the same cleanup — see the commit that follows this Phase 0.

android/app/src/main/java/com/wzp/debug/DebugReporter.kt

@@ -46,6 +46,14 @@ class DebugReporter(private val context: Context) {
             val zipFile = File(context.cacheDir, "wzp_debug_${timestamp}.zip")
 
             ZipOutputStream(BufferedOutputStream(FileOutputStream(zipFile))).use { zos ->
+                // Phase 4: extract DRED / classical PLC counters from the
+                // stats JSON so they're visible in the meta preamble at a
+                // glance, not buried in the trailing JSON dump.
+                val dredReconstructions = extractLongField(finalStatsJson, "dred_reconstructions")
+                val classicalPlc = extractLongField(finalStatsJson, "classical_plc_invocations")
+                val framesDecoded = extractLongField(finalStatsJson, "frames_decoded")
+                val fecRecovered = extractLongField(finalStatsJson, "fec_recovered")
+
                 // 1. Call metadata
                 val meta = buildString {
                     appendLine("=== WZ Phone Debug Report ===")
@@ -58,6 +66,18 @@ class DebugReporter(private val context: Context) {
                     appendLine("Device: ${android.os.Build.MANUFACTURER} ${android.os.Build.MODEL}")
                     appendLine("Android: ${android.os.Build.VERSION.RELEASE} (API ${android.os.Build.VERSION.SDK_INT})")
                     appendLine()
+                    appendLine("=== Loss Recovery ===")
+                    appendLine("Frames decoded:           $framesDecoded")
+                    appendLine("DRED reconstructions:     $dredReconstructions (Opus neural recovery)")
+                    appendLine("Classical PLC:            $classicalPlc (fallback)")
+                    appendLine("RaptorQ FEC recovered:    $fecRecovered (Codec2 only)")
+                    if (framesDecoded > 0) {
+                        val dredPct = 100.0 * dredReconstructions / framesDecoded
+                        val plcPct = 100.0 * classicalPlc / framesDecoded
+                        appendLine("DRED rate:                ${"%.2f".format(dredPct)}%")
+                        appendLine("Classical PLC rate:       ${"%.2f".format(plcPct)}%")
+                    }
+                    appendLine()
                     appendLine("=== Final Stats ===")
                     appendLine(finalStatsJson)
                 }
@@ -195,4 +215,28 @@ class DebugReporter(private val context: Context) {
         FileInputStream(file).use { it.copyTo(zos) }
         zos.closeEntry()
     }
+
+    /**
+     * Tiny JSON field extractor — pulls an integer value for a top-level
+     * field like `"dred_reconstructions":42`. We don't want to pull in a
+     * full JSON parser just for the debug preamble, and the CallStats
+     * output is a flat record with well-known field names.
+     *
+     * Returns 0 if the field is missing or unparseable.
+     */
+    private fun extractLongField(json: String, field: String): Long {
+        val key = "\"$field\":"
+        val idx = json.indexOf(key)
+        if (idx < 0) return 0
+        var i = idx + key.length
+        // Skip whitespace
+        while (i < json.length && json[i].isWhitespace()) i++
+        val start = i
+        while (i < json.length && (json[i].isDigit() || json[i] == '-')) i++
+        return try {
+            json.substring(start, i).toLong()
+        } catch (_: NumberFormatException) {
+            0
+        }
+    }
 }

android/app/src/main/java/com/wzp/engine/WzpEngine.kt

@@ -96,6 +96,17 @@ class WzpEngine(private val callback: WzpCallback) {
         if (nativeHandle != 0L) nativeForceProfile(nativeHandle, profile)
     }
 
+    /**
+     * Signal a network transport change (e.g. WiFi → LTE handoff).
+     *
+     * @param networkType matches Rust `NetworkContext` ordinals:
+     *   0=WiFi, 1=LTE, 2=5G, 3=3G, 4=Unknown, 5=None
+     * @param bandwidthKbps reported downstream bandwidth in kbps
+     */
+    fun onNetworkChanged(networkType: Int, bandwidthKbps: Int) {
+        if (nativeHandle != 0L) nativeOnNetworkChanged(nativeHandle, networkType, bandwidthKbps)
+    }
+
     /** Destroy the native engine and free all resources. The instance must not be reused. */
     @Synchronized
     fun destroy() {
@@ -163,6 +174,7 @@ class WzpEngine(private val callback: WzpCallback) {
     private external fun nativeStartSignaling(handle: Long, relay: String, seed: String, token: String, alias: String): Int
     private external fun nativePlaceCall(handle: Long, targetFp: String): Int
     private external fun nativeAnswerCall(handle: Long, callId: String, mode: Int): Int
+    private external fun nativeOnNetworkChanged(handle: Long, networkType: Int, bandwidthKbps: Int)
 
     /**
      * Ping a relay server. Requires engine to be initialized.

android/app/src/main/java/com/wzp/net/NetworkMonitor.kt

@@ -0,0 +1,141 @@
+package com.wzp.net
+
+import android.content.Context
+import android.net.ConnectivityManager
+import android.net.Network
+import android.net.NetworkCapabilities
+import android.net.NetworkRequest
+import android.os.Handler
+import android.os.Looper
+
+/**
+ * Monitors network connectivity changes via [ConnectivityManager.NetworkCallback]
+ * and classifies the active transport (WiFi, LTE, 5G, 3G).
+ *
+ * Callbacks fire on the main looper so callers can safely update UI state or
+ * dispatch to a native engine from any callback.
+ *
+ * Usage:
+ * 1. Set [onNetworkChanged] to receive `(type: Int, downlinkKbps: Int)` events
+ * 2. Optionally set [onIpChanged] for IP address change events (mid-call ICE refresh)
+ * 3. Call [register] when the call starts
+ * 4. Call [unregister] when the call ends
+ */
+class NetworkMonitor(context: Context) {
+
+    private val cm = context.getSystemService(Context.CONNECTIVITY_SERVICE) as ConnectivityManager
+    private val mainHandler = Handler(Looper.getMainLooper())
+
+    /**
+     * Called when the network transport type or bandwidth changes.
+     * `type` constants match the Rust `NetworkContext` enum ordinals.
+     */
+    var onNetworkChanged: ((type: Int, downlinkKbps: Int) -> Unit)? = null
+
+    /**
+     * Called when the device's IP address changes (link properties changed).
+     * Useful for triggering mid-call ICE candidate re-gathering.
+     */
+    var onIpChanged: (() -> Unit)? = null
+
+    // Track the last emitted type to avoid redundant callbacks
+    @Volatile
+    private var lastEmittedType: Int = TYPE_UNKNOWN
+
+    private val callback = object : ConnectivityManager.NetworkCallback() {
+        override fun onAvailable(network: Network) {
+            classifyAndEmit(network)
+        }
+
+        override fun onCapabilitiesChanged(network: Network, caps: NetworkCapabilities) {
+            classifyFromCaps(caps)
+        }
+
+        override fun onLinkPropertiesChanged(
+            network: Network,
+            linkProperties: android.net.LinkProperties
+        ) {
+            // IP address may have changed — notify for ICE refresh
+            onIpChanged?.invoke()
+            // Also re-classify in case the transport changed simultaneously
+            classifyAndEmit(network)
+        }
+
+        override fun onLost(network: Network) {
+            lastEmittedType = TYPE_NONE
+            onNetworkChanged?.invoke(TYPE_NONE, 0)
+        }
+    }
+
+    // -- Public API -----------------------------------------------------------
+
+    /** Register the network callback. Call when a call starts. */
+    fun register() {
+        val request = NetworkRequest.Builder()
+            .addCapability(NetworkCapabilities.NET_CAPABILITY_INTERNET)
+            .build()
+        cm.registerNetworkCallback(request, callback, mainHandler)
+    }
+
+    /** Unregister the network callback. Call when the call ends. */
+    fun unregister() {
+        try {
+            cm.unregisterNetworkCallback(callback)
+        } catch (_: IllegalArgumentException) {
+            // Already unregistered — safe to ignore
+        }
+    }
+
+    // -- Classification -------------------------------------------------------
+
+    private fun classifyAndEmit(network: Network) {
+        val caps = cm.getNetworkCapabilities(network) ?: return
+        classifyFromCaps(caps)
+    }
+
+    private fun classifyFromCaps(caps: NetworkCapabilities) {
+        val type = when {
+            caps.hasTransport(NetworkCapabilities.TRANSPORT_WIFI) -> TYPE_WIFI
+            caps.hasTransport(NetworkCapabilities.TRANSPORT_ETHERNET) -> TYPE_WIFI // treat as WiFi
+            caps.hasTransport(NetworkCapabilities.TRANSPORT_CELLULAR) -> classifyCellular(caps)
+            else -> TYPE_UNKNOWN
+        }
+        val bw = caps.getLinkDownstreamBandwidthKbps()
+
+        // Deduplicate: only emit when the transport type actually changes
+        if (type != lastEmittedType) {
+            lastEmittedType = type
+            onNetworkChanged?.invoke(type, bw)
+        }
+    }
+
+    /**
+     * Approximate cellular generation from reported downstream bandwidth.
+     * This avoids requiring READ_PHONE_STATE permission (needed for
+     * TelephonyManager.getNetworkType on API 30+).
+     *
+     * Thresholds are conservative — carriers over-report bandwidth, so we
+     * classify based on what's actually usable for VoIP:
+     * - >= 100 Mbps → 5G NR
+     * - >= 10 Mbps  → LTE
+     * - < 10 Mbps   → 3G or worse
+     */
+    private fun classifyCellular(caps: NetworkCapabilities): Int {
+        val bw = caps.getLinkDownstreamBandwidthKbps()
+        return when {
+            bw >= 100_000 -> TYPE_CELLULAR_5G
+            bw >= 10_000 -> TYPE_CELLULAR_LTE
+            else -> TYPE_CELLULAR_3G
+        }
+    }
+
+    companion object {
+        /** Constants matching Rust `NetworkContext` enum ordinals. */
+        const val TYPE_WIFI = 0
+        const val TYPE_CELLULAR_LTE = 1
+        const val TYPE_CELLULAR_5G = 2
+        const val TYPE_CELLULAR_3G = 3
+        const val TYPE_UNKNOWN = 4
+        const val TYPE_NONE = 5
+    }
+}

android/app/src/main/java/com/wzp/ui/call/CallViewModel.kt

@@ -5,6 +5,7 @@ import android.util.Log
 import androidx.lifecycle.ViewModel
 import androidx.lifecycle.viewModelScope
 import com.wzp.audio.AudioPipeline
+import com.wzp.audio.AudioRoute
 import com.wzp.audio.AudioRouteManager
 import com.wzp.data.SettingsRepository
 import com.wzp.debug.DebugReporter
@@ -12,6 +13,7 @@ import com.wzp.engine.CallStats
 import com.wzp.service.CallService
 import com.wzp.engine.WzpCallback
 import com.wzp.engine.WzpEngine
+import com.wzp.net.NetworkMonitor
 import kotlinx.coroutines.Dispatchers
 import kotlinx.coroutines.Job
 import kotlinx.coroutines.delay
@@ -43,6 +45,7 @@ class CallViewModel : ViewModel(), WzpCallback {
     private var engineInitialized = false
     private var audioPipeline: AudioPipeline? = null
     private var audioRouteManager: AudioRouteManager? = null
+    private var networkMonitor: NetworkMonitor? = null
     private var audioStarted = false
     private var appContext: Context? = null
     private var settings: SettingsRepository? = null
@@ -60,6 +63,9 @@ class CallViewModel : ViewModel(), WzpCallback {
     private val _isSpeaker = MutableStateFlow(false)
     val isSpeaker: StateFlow<Boolean> = _isSpeaker.asStateFlow()
 
+    private val _audioRoute = MutableStateFlow(AudioRoute.EARPIECE)
+    val audioRoute: StateFlow<AudioRoute> = _audioRoute.asStateFlow()
+
     private val _stats = MutableStateFlow(CallStats())
     val stats: StateFlow<CallStats> = _stats.asStateFlow()
 
@@ -226,7 +232,19 @@ class CallViewModel : ViewModel(), WzpCallback {
             audioPipeline = AudioPipeline(appCtx)
         }
         if (audioRouteManager == null) {
-            audioRouteManager = AudioRouteManager(appCtx)
+            audioRouteManager = AudioRouteManager(appCtx).also { arm ->
+                arm.onRouteChanged = { route ->
+                    _audioRoute.value = route
+                    _isSpeaker.value = (route == AudioRoute.SPEAKER)
+                }
+            }
+        }
+        if (networkMonitor == null) {
+            networkMonitor = NetworkMonitor(appCtx).also { nm ->
+                nm.onNetworkChanged = { type, bw ->
+                    engine?.onNetworkChanged(type, bw)
+                }
+            }
         }
         if (debugReporter == null) {
             debugReporter = DebugReporter(appCtx)
@@ -607,6 +625,27 @@ class CallViewModel : ViewModel(), WzpCallback {
         audioRouteManager?.setSpeaker(newSpeaker)
     }
 
+    /** Cycle audio output: Earpiece → Speaker → Bluetooth (if available) → Earpiece. */
+    fun cycleAudioRoute() {
+        val routes = audioRouteManager?.availableRoutes() ?: return
+        val currentIdx = routes.indexOf(_audioRoute.value)
+        val next = routes[(currentIdx + 1) % routes.size]
+        when (next) {
+            AudioRoute.EARPIECE -> {
+                audioRouteManager?.setBluetoothSco(false)
+                audioRouteManager?.setSpeaker(false)
+            }
+            AudioRoute.SPEAKER -> {
+                audioRouteManager?.setSpeaker(true)
+            }
+            AudioRoute.BLUETOOTH -> {
+                audioRouteManager?.setBluetoothSco(true)
+            }
+        }
+        _audioRoute.value = next
+        _isSpeaker.value = (next == AudioRoute.SPEAKER)
+    }
+
     fun clearError() { _errorMessage.value = null }
 
     fun sendDebugReport() {
@@ -661,6 +700,7 @@ class CallViewModel : ViewModel(), WzpCallback {
             it.start(e)
         }
         audioRouteManager?.register()
+        networkMonitor?.register()
         audioStarted = true
     }
 
@@ -668,8 +708,10 @@ class CallViewModel : ViewModel(), WzpCallback {
         if (!audioStarted) return
         audioPipeline?.stop()    // sets running=false; DON'T null — teardown needs awaitDrain()
         audioRouteManager?.unregister()
+        networkMonitor?.unregister()
         audioRouteManager?.setSpeaker(false)
         _isSpeaker.value = false
+        _audioRoute.value = AudioRoute.EARPIECE
         audioStarted = false
     }
 

android/app/src/main/java/com/wzp/ui/call/InCallScreen.kt

@@ -2,7 +2,6 @@ package com.wzp.ui.call
 
 import androidx.compose.foundation.background
 import androidx.compose.foundation.clickable
-import androidx.compose.ui.text.style.TextAlign
 import androidx.compose.foundation.layout.Arrangement
 import androidx.compose.foundation.layout.Box
 import androidx.compose.foundation.layout.Column
@@ -50,6 +49,7 @@ import androidx.compose.ui.text.font.FontWeight
 import androidx.compose.ui.text.style.TextAlign
 import androidx.compose.ui.unit.dp
 import androidx.compose.ui.unit.sp
+import com.wzp.audio.AudioRoute
 import com.wzp.engine.CallStats
 import com.wzp.ui.components.CopyableFingerprint
 import com.wzp.ui.components.Identicon
@@ -75,6 +75,7 @@ fun InCallScreen(
     val callState by viewModel.callState.collectAsState()
     val isMuted by viewModel.isMuted.collectAsState()
     val isSpeaker by viewModel.isSpeaker.collectAsState()
+    val audioRoute by viewModel.audioRoute.collectAsState()
     val stats by viewModel.stats.collectAsState()
     val qualityTier by viewModel.qualityTier.collectAsState()
     val errorMessage by viewModel.errorMessage.collectAsState()
@@ -622,12 +623,12 @@ fun InCallScreen(
 
                 Spacer(modifier = Modifier.height(16.dp))
 
-                // Controls: Mic / End / Spk
+                // Controls: Mic / End / Route (Ear/Spk/BT)
                 ControlRow(
                     isMuted = isMuted,
-                    isSpeaker = isSpeaker,
+                    audioRoute = audioRoute,
                     onToggleMute = viewModel::toggleMute,
-                    onToggleSpeaker = viewModel::toggleSpeaker,
+                    onCycleRoute = viewModel::cycleAudioRoute,
                     onHangUp = { viewModel.stopCall() }
                 )
 
@@ -916,9 +917,9 @@ private fun AudioLevelBar(audioLevel: Int) {
 @Composable
 private fun ControlRow(
     isMuted: Boolean,
-    isSpeaker: Boolean,
+    audioRoute: AudioRoute,
     onToggleMute: () -> Unit,
-    onToggleSpeaker: () -> Unit,
+    onCycleRoute: () -> Unit,
     onHangUp: () -> Unit
 ) {
     Row(
@@ -960,22 +961,28 @@ private fun ControlRow(
             Text("End", style = MaterialTheme.typography.titleMedium.copy(fontWeight = FontWeight.Bold))
         }
 
-        // Speaker
+        // Audio route: cycles Earpiece → Speaker → Bluetooth (when available)
         FilledTonalIconButton(
-            onClick = onToggleSpeaker,
+            onClick = onCycleRoute,
             modifier = Modifier.size(56.dp),
-            colors = if (isSpeaker) {
-                IconButtonDefaults.filledTonalIconButtonColors(
+            colors = when (audioRoute) {
+                AudioRoute.SPEAKER -> IconButtonDefaults.filledTonalIconButtonColors(
                     containerColor = Color(0xFF0F3460), contentColor = Color.White
                 )
-            } else {
-                IconButtonDefaults.filledTonalIconButtonColors(
+                AudioRoute.BLUETOOTH -> IconButtonDefaults.filledTonalIconButtonColors(
+                    containerColor = Color(0xFF2563EB), contentColor = Color.White
+                )
+                else -> IconButtonDefaults.filledTonalIconButtonColors(
                     containerColor = DarkSurface2, contentColor = Color.White
                 )
             }
         ) {
             Text(
-                text = if (isSpeaker) "Spk\nOn" else "Spk",
+                text = when (audioRoute) {
+                    AudioRoute.EARPIECE -> "Ear"
+                    AudioRoute.SPEAKER -> "Spk"
+                    AudioRoute.BLUETOOTH -> "BT"
+                },
                 textAlign = TextAlign.Center,
                 style = MaterialTheme.typography.labelSmall,
                 lineHeight = 12.sp

crates/wzp-android/src/engine.rs

@@ -14,8 +14,10 @@ use std::sync::{Arc, Mutex};
 use std::time::Instant;
 
 use bytes::Bytes;
-use tracing::{error, info, warn};
+use tracing::{debug, error, info, warn};
+use wzp_codec::AdaptiveDecoder;
 use wzp_codec::agc::AutoGainControl;
+use wzp_codec::dred_ffi::{DredDecoderHandle, DredState};
 use wzp_crypto::{KeyExchange, WarzoneKeyExchange};
 use wzp_fec::{RaptorQFecDecoder, RaptorQFecEncoder};
 use wzp_proto::{
@@ -97,6 +99,9 @@ pub(crate) struct EngineState {
     /// QUIC transport handle — stored so stop_call() can close it immediately,
     /// triggering relay-side leave + RoomUpdate broadcast.
     pub quic_transport: Mutex<Option<Arc<wzp_transport::QuinnTransport>>>,
+    /// Network type from Android ConnectivityManager, polled by recv task.
+    /// 0xFF = no change pending; 0-5 = NetworkContext ordinal.
+    pub pending_network_type: AtomicU8,
 }
 
 pub struct WzpEngine {
@@ -118,6 +123,7 @@ impl WzpEngine {
             playout_ring: AudioRing::new(),
             audio_level_rms: AtomicU32::new(0),
             quic_transport: Mutex::new(None),
+            pending_network_type: AtomicU8::new(PROFILE_NO_CHANGE),
         });
         Self {
             state,
@@ -340,7 +346,7 @@ impl WzpEngine {
                     Ok(Some(SignalMessage::DirectCallAnswer { call_id, accept_mode, .. })) => {
                         info!(call_id = %call_id, mode = ?accept_mode, "signal: call answered");
                     }
-                    Ok(Some(SignalMessage::CallSetup { call_id, room, relay_addr })) => {
+                    Ok(Some(SignalMessage::CallSetup { call_id, room, relay_addr, .. })) => {
                         info!(call_id = %call_id, room = %room, relay = %relay_addr, "signal: call setup");
                         // Connect to media room via the existing start_call mechanism
                         // Store the room info so Kotlin can call startCall with it
@@ -349,7 +355,7 @@ impl WzpEngine {
                         // Store call setup info for Kotlin to pick up
                         stats.incoming_call_id = Some(format!("{relay_addr}|{room}"));
                     }
-                    Ok(Some(SignalMessage::Hangup { reason })) => {
+                    Ok(Some(SignalMessage::Hangup { reason, .. })) => {
                         info!(reason = ?reason, "signal: call ended by remote");
                         let mut stats = signal_state.stats.lock().unwrap();
                         stats.state = crate::stats::CallState::Closed;
@@ -402,6 +408,13 @@ impl WzpEngine {
 
     pub fn force_profile(&self, _profile: QualityProfile) {}
 
+    /// Signal a network transport change from Android ConnectivityManager.
+    /// Stores the type atomically; the recv task polls it on each packet.
+    pub fn on_network_changed(&self, network_type: u8, bandwidth_kbps: u32) {
+        info!(network_type, bandwidth_kbps, "on_network_changed");
+        self.state.pending_network_type.store(network_type, Ordering::Release);
+    }
+
     pub fn get_stats(&self) -> CallStats {
         let mut stats = self.state.stats.lock().unwrap().clone();
         if let Some(start) = self.call_start {
@@ -530,9 +543,12 @@ async fn run_call(
         stats.state = CallState::Active;
     }
 
-    // Initialize codec (Opus or Codec2 based on profile)
+    // Initialize codec (Opus or Codec2 based on profile).
+    // Phase 3c: decoder is a concrete AdaptiveDecoder (not Box<dyn
+    // AudioDecoder>) so the recv task can call reconstruct_from_dred on
+    // gaps detected via sequence tracking.
     let mut encoder = wzp_codec::create_encoder(profile);
-    let mut decoder = wzp_codec::create_decoder(profile);
+    let mut decoder = AdaptiveDecoder::new(profile).expect("failed to create adaptive decoder");
 
     // Initialize FEC encoder/decoder
     let mut fec_enc = wzp_fec::create_encoder(&profile);
@@ -665,6 +681,19 @@ async fn run_call(
             t_opus_us += t0.elapsed().as_micros() as u64;
             let encoded = &encode_buf[..encoded_len];
 
+            // Phase 2: Opus tiers bypass RaptorQ (DRED handles loss recovery
+            // at the codec layer). Codec2 tiers keep RaptorQ unchanged.
+            let is_opus = current_profile.codec.is_opus();
+            let (hdr_fec_block, hdr_fec_symbol, hdr_fec_ratio) = if is_opus {
+                (0u8, 0u8, 0u8)
+            } else {
+                (
+                    block_id,
+                    frame_in_block,
+                    MediaHeader::encode_fec_ratio(current_profile.fec_ratio),
+                )
+            };
+
             // Build source packet
             let s = seq.fetch_add(1, Ordering::Relaxed);
             let t = ts.fetch_add(frame_samples as u32, Ordering::Relaxed);
@@ -675,11 +704,11 @@ async fn run_call(
                     is_repair: false,
                     codec_id: current_profile.codec,
                     has_quality_report: false,
-                    fec_ratio_encoded: MediaHeader::encode_fec_ratio(current_profile.fec_ratio),
+                    fec_ratio_encoded: hdr_fec_ratio,
                     seq: s,
                     timestamp: t,
-                    fec_block: block_id,
-                    fec_symbol: frame_in_block,
+                    fec_block: hdr_fec_block,
+                    fec_symbol: hdr_fec_symbol,
                     reserved: 0,
                     csrc_count: 0,
                 },
@@ -709,63 +738,66 @@ async fn run_call(
             t_send_us += t0.elapsed().as_micros() as u64;
             frames_sent += 1;
 
-            // Feed encoded frame to FEC encoder
+            // Codec2-only: feed RaptorQ and emit repair packets when the
+            // block is full. Opus tiers skip this entire block — DRED
+            // (enabled in Phase 1) provides codec-layer loss recovery.
             let t0 = Instant::now();
-            if let Err(e) = fec_enc.add_source_symbol(encoded) {
-                warn!("fec add_source error: {e}");
-            }
-            frame_in_block += 1;
+            if !is_opus {
+                if let Err(e) = fec_enc.add_source_symbol(encoded) {
+                    warn!("fec add_source error: {e}");
+                }
+                frame_in_block += 1;
 
-            // When block is full, generate repair packets
-            if frame_in_block >= current_profile.frames_per_block {
-                match fec_enc.generate_repair(current_profile.fec_ratio) {
-                    Ok(repairs) => {
-                        let repair_count = repairs.len();
-                        for (sym_idx, repair_data) in repairs {
-                            let rs = seq.fetch_add(1, Ordering::Relaxed);
-                            let repair_pkt = MediaPacket {
-                                header: MediaHeader {
-                                    version: 0,
-                                    is_repair: true,
-                                    codec_id: current_profile.codec,
-                                    has_quality_report: false,
-                                    fec_ratio_encoded: MediaHeader::encode_fec_ratio(
-                                        current_profile.fec_ratio,
-                                    ),
-                                    seq: rs,
-                                    timestamp: t,
-                                    fec_block: block_id,
-                                    fec_symbol: sym_idx,
-                                    reserved: 0,
-                                    csrc_count: 0,
-                                },
-                                payload: Bytes::from(repair_data),
-                                quality_report: None,
-                            };
-                            // Drop repair packets on error — never break
-                            if let Err(_e) = transport.send_media(&repair_pkt).await {
-                                send_errors += 1;
-                                frames_dropped += 1;
-                                // Don't log every repair failure — source error log covers it
+                if frame_in_block >= current_profile.frames_per_block {
+                    match fec_enc.generate_repair(current_profile.fec_ratio) {
+                        Ok(repairs) => {
+                            let repair_count = repairs.len();
+                            for (sym_idx, repair_data) in repairs {
+                                let rs = seq.fetch_add(1, Ordering::Relaxed);
+                                let repair_pkt = MediaPacket {
+                                    header: MediaHeader {
+                                        version: 0,
+                                        is_repair: true,
+                                        codec_id: current_profile.codec,
+                                        has_quality_report: false,
+                                        fec_ratio_encoded: MediaHeader::encode_fec_ratio(
+                                            current_profile.fec_ratio,
+                                        ),
+                                        seq: rs,
+                                        timestamp: t,
+                                        fec_block: block_id,
+                                        fec_symbol: sym_idx,
+                                        reserved: 0,
+                                        csrc_count: 0,
+                                    },
+                                    payload: Bytes::from(repair_data),
+                                    quality_report: None,
+                                };
+                                // Drop repair packets on error — never break
+                                if let Err(_e) = transport.send_media(&repair_pkt).await {
+                                    send_errors += 1;
+                                    frames_dropped += 1;
+                                    // Don't log every repair failure — source error log covers it
+                                }
+                            }
+                            if repair_count > 0 && (block_id % 50 == 0 || block_id == 0) {
+                                info!(
+                                    block_id,
+                                    repair_count,
+                                    fec_ratio = current_profile.fec_ratio,
+                                    "FEC block complete"
+                                );
                             }
                         }
-                        if repair_count > 0 && (block_id % 50 == 0 || block_id == 0) {
-                            info!(
-                                block_id,
-                                repair_count,
-                                fec_ratio = current_profile.fec_ratio,
-                                "FEC block complete"
-                            );
+                        Err(e) => {
+                            warn!("fec generate_repair error: {e}");
                         }
                     }
-                    Err(e) => {
-                        warn!("fec generate_repair error: {e}");
-                    }
-                }
 
-                let _ = fec_enc.finalize_block();
-                block_id = block_id.wrapping_add(1);
-                frame_in_block = 0;
+                    let _ = fec_enc.finalize_block();
+                    block_id = block_id.wrapping_add(1);
+                    frame_in_block = 0;
+                }
             }
             t_fec_us += t0.elapsed().as_micros() as u64;
             t_frames += 1;
@@ -808,7 +840,27 @@ async fn run_call(
         let mut last_stats_log = Instant::now();
         let mut quality_ctrl = AdaptiveQualityController::new();
         let mut last_peer_codec: Option<CodecId> = None;
-        info!("recv task started (Opus + RaptorQ FEC)");
+
+        // Phase 3c: DRED reconstruction state. Unlike the desktop
+        // CallDecoder (which sits behind a jitter buffer that emits
+        // Missing signals), engine.rs reads packets directly from the
+        // transport and decodes straight into the playout ring. Gap
+        // detection is therefore done via sequence-number tracking:
+        // when a packet arrives with seq > expected_seq, the frames in
+        // between are missing and we attempt to reconstruct them via
+        // DRED before decoding the newly-arrived packet.
+        let mut dred_decoder =
+            DredDecoderHandle::new().expect("opus_dred_decoder_create failed");
+        let mut dred_parse_scratch =
+            DredState::new().expect("opus_dred_alloc failed (scratch)");
+        let mut last_good_dred =
+            DredState::new().expect("opus_dred_alloc failed (good state)");
+        let mut last_good_dred_seq: Option<u16> = None;
+        let mut expected_seq: Option<u16> = None;
+        let mut dred_reconstructions: u64 = 0;
+        let mut classical_plc_invocations: u64 = 0;
+
+        info!("recv task started (Opus + DRED + Codec2/RaptorQ)");
         loop {
             if !state.running.load(Ordering::Relaxed) {
                 break;
@@ -830,6 +882,23 @@ async fn run_call(
                         );
                     }
 
+                    // Check for network transport change from ConnectivityManager
+                    {
+                        let net = state.pending_network_type.swap(PROFILE_NO_CHANGE, Ordering::Acquire);
+                        if net != PROFILE_NO_CHANGE {
+                            use wzp_proto::NetworkContext;
+                            let ctx = match net {
+                                0 => NetworkContext::WiFi,
+                                1 => NetworkContext::CellularLte,
+                                2 => NetworkContext::Cellular5g,
+                                3 => NetworkContext::Cellular3g,
+                                _ => NetworkContext::Unknown,
+                            };
+                            quality_ctrl.signal_network_change(ctx);
+                            info!(?ctx, "quality controller: network context updated");
+                        }
+                    }
+
                     // Adaptive quality: ingest quality reports from relay
                     if auto_profile {
                         if let Some(ref qr) = pkt.quality_report {
@@ -850,14 +919,21 @@ async fn run_call(
                     let is_repair = pkt.header.is_repair;
                     let pkt_block = pkt.header.fec_block;
                     let pkt_symbol = pkt.header.fec_symbol;
+                    let pkt_is_opus = pkt.header.codec_id.is_opus();
 
-                    // Feed every packet (source + repair) to FEC decoder
-                    let _ = fec_dec.add_symbol(
-                        pkt_block,
-                        pkt_symbol,
-                        is_repair,
-                        &pkt.payload,
-                    );
+                    // Phase 2: Opus packets bypass RaptorQ entirely — DRED
+                    // (enabled Phase 1) handles codec-layer loss recovery,
+                    // and feeding these symbols into the RaptorQ decoder
+                    // would accumulate block_id=0 duplicates that never
+                    // decode. Codec2 packets still feed RaptorQ.
+                    if !pkt_is_opus {
+                        let _ = fec_dec.add_symbol(
+                            pkt_block,
+                            pkt_symbol,
+                            is_repair,
+                            &pkt.payload,
+                        );
+                    }
 
                     // Source packets: decode directly
                     if !is_repair && pkt.header.codec_id != CodecId::ComfortNoise {
@@ -880,6 +956,13 @@ async fn run_call(
                             };
                             info!(from = ?decoder.codec_id(), to = ?pkt.header.codec_id, "recv: switching decoder");
                             let _ = decoder.set_profile(switch_profile);
+                            // Profile switch invalidates the cached DRED
+                            // state because samples_available is measured
+                            // in the old profile's sample rate. Reset the
+                            // tracking so we don't try to reconstruct with
+                            // stale offsets.
+                            last_good_dred_seq = None;
+                            expected_seq = None;
                         }
                         // Track peer codec for UI display
                         if last_peer_codec != Some(pkt.header.codec_id) {
@@ -888,6 +971,109 @@ async fn run_call(
                                 stats.peer_codec = format!("{:?}", pkt.header.codec_id);
                             }
                         }
+
+                        // Phase 3c: Opus path — parse DRED state out of
+                        // the current packet FIRST so last_good_dred
+                        // reflects the freshest available reconstruction
+                        // source, then attempt gap recovery against it
+                        // BEFORE decoding this packet's audio. Ordering
+                        // matters because the playout ring is FIFO — gap
+                        // samples must be written before this packet's
+                        // samples, which come next.
+                        if pkt_is_opus {
+                            // Update DRED state from the current packet.
+                            match dred_decoder.parse_into(&mut dred_parse_scratch, &pkt.payload) {
+                                Ok(available) if available > 0 => {
+                                    std::mem::swap(
+                                        &mut dred_parse_scratch,
+                                        &mut last_good_dred,
+                                    );
+                                    last_good_dred_seq = Some(pkt.header.seq);
+                                }
+                                Ok(_) => {
+                                    // Packet carried no DRED — keep cached state.
+                                }
+                                Err(e) => {
+                                    debug!("DRED parse error (ignored): {e}");
+                                }
+                            }
+
+                            // Detect and fill gap from last-expected to this packet.
+                            const MAX_GAP_FRAMES: u16 = 16;
+                            if let Some(expected) = expected_seq {
+                                let gap = pkt.header.seq.wrapping_sub(expected);
+                                if gap > 0 && gap <= MAX_GAP_FRAMES {
+                                    let current_profile_frame_samples =
+                                        (48_000 * profile.frame_duration_ms as i32) / 1000;
+                                    let available = last_good_dred.samples_available();
+                                    let pcm_slice_len =
+                                        current_profile_frame_samples as usize;
+
+                                    for gap_idx in 0..gap {
+                                        let missing_seq = expected.wrapping_add(gap_idx);
+                                        // Offset from the DRED anchor (last_good_dred_seq)
+                                        // back to the missing seq, in samples. Skip if
+                                        // the anchor is not ahead of missing (defensive).
+                                        let offset_samples = match last_good_dred_seq {
+                                            Some(anchor) => {
+                                                let delta = anchor.wrapping_sub(missing_seq);
+                                                if delta == 0 || delta > MAX_GAP_FRAMES {
+                                                    -1 // skip DRED, use PLC
+                                                } else {
+                                                    delta as i32 * current_profile_frame_samples
+                                                }
+                                            }
+                                            None => -1,
+                                        };
+
+                                        let reconstructed = if offset_samples > 0
+                                            && offset_samples <= available
+                                        {
+                                            decoder
+                                                .reconstruct_from_dred(
+                                                    &last_good_dred,
+                                                    offset_samples,
+                                                    &mut decode_buf[..pcm_slice_len],
+                                                )
+                                                .ok()
+                                        } else {
+                                            None
+                                        };
+
+                                        match reconstructed {
+                                            Some(samples) => {
+                                                playout_agc.process_frame(
+                                                    &mut decode_buf[..samples],
+                                                );
+                                                state
+                                                    .playout_ring
+                                                    .write(&decode_buf[..samples]);
+                                                dred_reconstructions += 1;
+                                                frames_decoded += 1;
+                                            }
+                                            None => {
+                                                // Fall through to classical PLC.
+                                                if let Ok(samples) =
+                                                    decoder.decode_lost(&mut decode_buf)
+                                                {
+                                                    playout_agc
+                                                        .process_frame(&mut decode_buf[..samples]);
+                                                    state
+                                                        .playout_ring
+                                                        .write(&decode_buf[..samples]);
+                                                    classical_plc_invocations += 1;
+                                                    frames_decoded += 1;
+                                                }
+                                            }
+                                        }
+                                    }
+                                }
+                            }
+
+                            // Advance the expected-seq tracker for the next arrival.
+                            expected_seq = Some(pkt.header.seq.wrapping_add(1));
+                        }
+
                         match decoder.decode(&pkt.payload, &mut decode_buf) {
                             Ok(samples) => {
                                 playout_agc.process_frame(&mut decode_buf[..samples]);
@@ -899,32 +1085,44 @@ async fn run_call(
                                 if let Ok(samples) = decoder.decode_lost(&mut decode_buf) {
                                     playout_agc.process_frame(&mut decode_buf[..samples]);
                                     state.playout_ring.write(&decode_buf[..samples]);
+                                    // This is a decode-error fallback (not a
+                                    // detected gap), so count it as PLC.
+                                    classical_plc_invocations += 1;
                                 }
                             }
                         }
                     }
 
-                    // Try FEC recovery
-                    if let Ok(Some(recovered_frames)) = fec_dec.try_decode(pkt_block) {
-                        fec_recovered += recovered_frames.len() as u64;
-                        if fec_recovered % 50 == 1 {
-                            info!(
-                                fec_recovered,
-                                block = pkt_block,
-                                frames = recovered_frames.len(),
-                                "FEC block recovered"
-                            );
+                    // Codec2-only: try FEC recovery and expire old blocks.
+                    // Opus packets skip both — the Phase 2 Opus path has no
+                    // RaptorQ state to query or clean up. The `fec_recovered`
+                    // counter is now effectively Codec2-only, which is
+                    // correct because DRED reconstructions will be counted
+                    // separately once Phase 3 lands (new telemetry field).
+                    if !pkt_is_opus {
+                        if let Ok(Some(recovered_frames)) = fec_dec.try_decode(pkt_block) {
+                            fec_recovered += recovered_frames.len() as u64;
+                            if fec_recovered % 50 == 1 {
+                                info!(
+                                    fec_recovered,
+                                    block = pkt_block,
+                                    frames = recovered_frames.len(),
+                                    "FEC block recovered"
+                                );
+                            }
                         }
-                    }
 
-                    // Expire old blocks to prevent memory growth
-                    if pkt_block > 3 {
-                        fec_dec.expire_before(pkt_block.wrapping_sub(3));
+                        // Expire old blocks to prevent memory growth
+                        if pkt_block > 3 {
+                            fec_dec.expire_before(pkt_block.wrapping_sub(3));
+                        }
                     }
 
                     let mut stats = state.stats.lock().unwrap();
                     stats.frames_decoded = frames_decoded;
                     stats.fec_recovered = fec_recovered;
+                    stats.dred_reconstructions = dred_reconstructions;
+                    stats.classical_plc_invocations = classical_plc_invocations;
                     drop(stats);
 
                     // Periodic stats every 5 seconds
@@ -932,6 +1130,8 @@ async fn run_call(
                         info!(
                             frames_decoded,
                             fec_recovered,
+                            dred_reconstructions,
+                            classical_plc_invocations,
                             recv_errors,
                             max_recv_gap_ms,
                             playout_avail = state.playout_ring.available(),
@@ -1009,6 +1209,15 @@ async fn run_call(
                     stats.room_participant_count = count;
                     stats.room_participants = members;
                 }
+                Ok(Some(SignalMessage::QualityDirective { recommended_profile, reason })) => {
+                    let idx = profile_to_index(&recommended_profile);
+                    info!(
+                        codec = ?recommended_profile.codec,
+                        reason = reason.as_deref().unwrap_or(""),
+                        "relay quality directive: switching profile"
+                    );
+                    pending_profile_recv.store(idx, Ordering::Release);
+                }
                 Ok(Some(msg)) => {
                     info!("signal received: {:?}", std::mem::discriminant(&msg));
                 }

crates/wzp-android/src/jni_bridge.rs

@@ -222,6 +222,29 @@ pub unsafe extern "system" fn Java_com_wzp_engine_WzpEngine_nativeForceProfile(
     }));
 }
 
+/// Signal a network transport change from the Android ConnectivityManager.
+///
+/// `network_type` matches the Rust `NetworkContext` enum:
+///   0=WiFi, 1=CellularLte, 2=Cellular5g, 3=Cellular3g, 4=Unknown, 5=None
+///
+/// The engine forwards this to the `AdaptiveQualityController` which:
+/// - Preemptively downgrades one tier on WiFi→cellular
+/// - Activates a 10-second FEC boost
+/// - Uses faster downgrade thresholds on cellular
+#[unsafe(no_mangle)]
+pub unsafe extern "system" fn Java_com_wzp_engine_WzpEngine_nativeOnNetworkChanged(
+    _env: JNIEnv,
+    _class: JClass,
+    handle: jlong,
+    network_type: jint,
+    bandwidth_kbps: jint,
+) {
+    let _ = panic::catch_unwind(panic::AssertUnwindSafe(|| {
+        let h = unsafe { handle_ref(handle) };
+        h.engine.on_network_changed(network_type as u8, bandwidth_kbps as u32);
+    }));
+}
+
 /// Write captured PCM samples from Kotlin AudioRecord into the engine's capture ring.
 /// pcm is a Java short[] array.
 #[unsafe(no_mangle)]

crates/wzp-android/src/lib.rs

@@ -8,6 +8,19 @@
 //!
 //! On non-Android targets, the Oboe C++ layer compiles as a stub,
 //! allowing `cargo check` and unit tests on the host.
+//!
+//! ## Status
+//!
+//! **Dead code as of the Tauri mobile rewrite.** The legacy Kotlin+JNI
+//! Android app that consumed this crate was replaced by a Tauri 2.x
+//! Mobile app (see `desktop/src-tauri/src/engine.rs` for the live
+//! Android audio recv path and `crates/wzp-native/` for the Oboe
+//! bridge). We keep this crate in the workspace for reference and to
+//! preserve the commit history, but it is not built by any shipping
+//! target. Allow the accumulated leftover warnings so CI/workspace
+//! checks stay clean — any real cleanup should happen as part of
+//! removing the crate entirely, not piecemeal.
+#![allow(dead_code, unused_imports, unused_variables, unused_mut)]
 
 pub mod audio_android;
 pub mod audio_ring;

crates/wzp-android/src/stats.rs

@@ -58,8 +58,16 @@ pub struct CallStats {
     pub frames_decoded: u64,
     /// Number of playout underruns (buffer empty when audio needed).
     pub underruns: u64,
-    /// Frames recovered by FEC.
+    /// Frames recovered by RaptorQ FEC (Codec2 tiers only; Opus bypasses
+    /// RaptorQ per Phase 2).
     pub fec_recovered: u64,
+    /// Phase 3c: Opus frames reconstructed via DRED side-channel data.
+    /// Only increments on the Opus tiers; always zero for Codec2.
+    pub dred_reconstructions: u64,
+    /// Phase 3c: Opus frames filled via classical Opus PLC because no DRED
+    /// state covered the gap, plus any decode-error fallbacks. Codec2 loss
+    /// also increments this counter via the Codec2 PLC path.
+    pub classical_plc_invocations: u64,
     /// Playout ring overflow count (reader was lapped by writer).
     pub playout_overflows: u64,
     /// Playout ring underrun count (reader found empty buffer).

crates/wzp-client/Cargo.toml

@@ -21,17 +21,93 @@ anyhow = "1"
 serde = { workspace = true }
 serde_json = "1"
 chrono = "0.4"
+clap = { version = "4", features = ["derive"] }
+ratatui = "0.29"
+crossterm = "0.28"
 rustls = { version = "0.23", default-features = false, features = ["ring", "std"] }
 cpal = { version = "0.15", optional = true }
+libc = "0.2"
+# Phase 5.5 — LAN host-candidate ICE: enumerate local network
+# interface addresses for inclusion in DirectCallOffer/Answer so
+# peers on the same LAN can direct-connect without NAT hairpinning
+# through the WAN reflex addr (which many consumer NATs, including
+# MikroTik's default masquerade, don't support).
+if-addrs = "0.13"
+rand = { workspace = true }
+socket2 = "0.5"
+
+# coreaudio-rs is Apple-framework-only; gate it to macOS so enabling
+# the `vpio` feature from a non-macOS target builds cleanly instead of
+# pulling in a crate that can only link against Apple frameworks.
+[target.'cfg(target_os = "macos")'.dependencies]
+coreaudio-rs = { version = "0.11", optional = true }
+
+# Windows-only: direct WASAPI bindings for the `windows-aec` feature.
+# `windows` is Microsoft's official Rust COM bindings crate. We pull in
+# only the audio + COM subfeatures we need — the crate is organized as
+# a massive optional-feature tree, so enabling just these keeps compile
+# times reasonable (~5s for these features vs ~60s for the full crate).
+[target.'cfg(target_os = "windows")'.dependencies]
+windows = { version = "0.58", optional = true, features = [
+    "Win32_Foundation",
+    "Win32_Media_Audio",
+    "Win32_Security",
+    "Win32_System_Com",
+    "Win32_System_Com_StructuredStorage",
+    "Win32_System_Threading",
+    "Win32_System_Variant",
+] }
+
+# Linux-only: WebRTC AEC (Audio Processing Module) bindings for the
+# `linux-aec` feature. This is the 0.3.x line of the `tonarino/
+# webrtc-audio-processing` crate, which links against Debian's
+# `libwebrtc-audio-processing-dev` apt package (0.3-1+b1 on Bookworm).
+#
+# Note: we attempted the 2.x line with its `bundled` sub-feature first
+# (which would give us AEC3 instead of AEC2), but both the crates.io
+# tarball AND the upstream git `main` branch of webrtc-audio-processing-sys
+# 2.0.3 hit a `meson setup --reconfigure` bug where the build.rs passes
+# --reconfigure unconditionally even on first-run empty build dirs,
+# causing the bundled build to fail with "Directory does not contain a
+# valid build tree". The 0.x line doesn't use bundled mode and sidesteps
+# this entirely by linking the apt-provided library. AEC2 is older than
+# AEC3 but still the same algorithm family — this is what PulseAudio's
+# module-echo-cancel and PipeWire's filter-chain use by default on
+# current Debian-family distros.
+[target.'cfg(target_os = "linux")'.dependencies]
+webrtc-audio-processing = { version = "0.3", optional = true }
 
 [features]
 default = []
 audio = ["cpal"]
+# vpio enables coreaudio-rs but that dep is itself gated to macOS above,
+# so enabling this feature on Windows/Linux is a no-op (the audio_vpio
+# module is also #[cfg(target_os = "macos")] in lib.rs).
+vpio = ["dep:coreaudio-rs"]
+# windows-aec enables a direct WASAPI capture backend that opens the
+# microphone under AudioCategory_Communications, turning on Windows's
+# OS-level communications audio processing (AEC + noise suppression +
+# AGC). The `windows` dep is itself target-gated to Windows above, so
+# enabling this feature on non-Windows targets is a no-op (the
+# audio_wasapi module is also #[cfg(target_os = "windows")] in lib.rs).
+windows-aec = ["dep:windows"]
+# linux-aec enables a CPAL + WebRTC AEC3 capture/playback backend that
+# runs the WebRTC Audio Processing Module (same algo as Chrome / Zoom /
+# Teams) in-process, using the playback PCM as the reference signal for
+# echo cancellation. The webrtc-audio-processing dep is target-gated to
+# Linux above, so enabling this feature on non-Linux targets is a no-op
+# (the audio_linux_aec module is also #[cfg(target_os = "linux")] in
+# lib.rs).
+linux-aec = ["dep:webrtc-audio-processing"]
 
 [[bin]]
 name = "wzp-client"
 path = "src/cli.rs"
 
+[[bin]]
+name = "wzp-analyzer"
+path = "src/analyzer.rs"
+
 [[bin]]
 name = "wzp-bench"
 path = "src/bench_cli.rs"

crates/wzp-client/src/analyzer.rs

@@ -0,0 +1,952 @@
+//! WarzonePhone Protocol Analyzer — passive call quality observer.
+//!
+//! Joins a relay room as a passive participant (no media sent) and displays
+//! real-time per-participant quality metrics in a terminal UI.
+//!
+//! Usage:
+//!   wzp-analyzer 127.0.0.1:4433 --room test
+//!   wzp-analyzer 1.2.3.4:4433 --room test --capture session.wzp
+//!   wzp-analyzer 1.2.3.4:4433 --room test --no-tui --duration 60
+
+use std::io::Write;
+use std::sync::Arc;
+use std::time::{Duration, Instant};
+
+use clap::Parser;
+use tracing::info;
+
+use wzp_proto::{CodecId, MediaPacket, MediaTransport};
+
+// ---------------------------------------------------------------------------
+// CLI
+// ---------------------------------------------------------------------------
+
+/// WarzonePhone Protocol Analyzer — passive call quality observer
+#[derive(Parser)]
+#[command(name = "wzp-analyzer", version)]
+struct Args {
+    /// Relay address (host:port) — required for live mode, ignored with --replay
+    relay: Option<String>,
+
+    /// Room name to observe — required for live mode, ignored with --replay
+    #[arg(short, long)]
+    room: Option<String>,
+
+    /// Auth token for relay
+    #[arg(long)]
+    token: Option<String>,
+
+    /// Identity seed (64-char hex)
+    #[arg(long)]
+    seed: Option<String>,
+
+    /// Capture packets to file
+    #[arg(long)]
+    capture: Option<String>,
+
+    /// Auto-stop after N seconds
+    #[arg(long)]
+    duration: Option<u64>,
+
+    /// Disable TUI (print stats to stdout instead)
+    #[arg(long)]
+    no_tui: bool,
+
+    /// Replay a captured .wzp file (offline analysis)
+    #[arg(long)]
+    replay: Option<String>,
+
+    /// Generate HTML report (from live session or replay)
+    #[arg(long)]
+    html: Option<String>,
+
+    /// Session key hex for decrypting payloads (enables audio decode)
+    // TODO(#17): Audio decode requires session key + nonce context.
+    // In SFU mode, payloads are E2E encrypted. Decoding requires
+    // either: (a) session key from both endpoints, or (b) running
+    // the analyzer as a trusted participant with its own key exchange.
+    // For now, header-only analysis provides loss%, jitter, codec stats.
+    #[arg(long)]
+    key: Option<String>,
+}
+
+// ---------------------------------------------------------------------------
+// Per-participant statistics
+// ---------------------------------------------------------------------------
+
+struct ParticipantStats {
+    /// Stream identifier (index, assigned when we detect a new seq stream)
+    stream_id: usize,
+    /// Display name from RoomUpdate (if available)
+    alias: Option<String>,
+    /// Current codec
+    codec: CodecId,
+    /// Total packets received
+    packets: u64,
+    /// Detected lost packets (sequence gaps)
+    lost: u64,
+    /// Last seen sequence number
+    last_seq: u16,
+    /// Whether we've seen the first packet (for gap detection)
+    seq_initialized: bool,
+    /// EWMA jitter in ms
+    jitter_ms: f64,
+    /// Last packet arrival time
+    last_arrival: Option<Instant>,
+    /// Codec changes observed
+    codec_switches: u32,
+    /// First packet time
+    first_seen: Instant,
+    /// Last packet time
+    last_seen: Instant,
+}
+
+impl ParticipantStats {
+    fn new(id: usize, codec: CodecId) -> Self {
+        let now = Instant::now();
+        Self {
+            stream_id: id,
+            alias: None,
+            codec,
+            packets: 0,
+            lost: 0,
+            last_seq: 0,
+            seq_initialized: false,
+            jitter_ms: 0.0,
+            last_arrival: None,
+            codec_switches: 0,
+            first_seen: now,
+            last_seen: now,
+        }
+    }
+
+    fn ingest(&mut self, pkt: &MediaPacket, now: Instant) {
+        self.packets += 1;
+        self.last_seen = now;
+
+        // Codec switch detection
+        if pkt.header.codec_id != self.codec {
+            self.codec_switches += 1;
+            self.codec = pkt.header.codec_id;
+        }
+
+        // Loss detection from sequence gaps
+        if self.seq_initialized {
+            let expected = self.last_seq.wrapping_add(1);
+            let gap = pkt.header.seq.wrapping_sub(expected);
+            if gap > 0 && gap < 100 {
+                self.lost += gap as u64;
+            }
+        }
+        self.last_seq = pkt.header.seq;
+        self.seq_initialized = true;
+
+        // Jitter (inter-arrival time variance, EWMA)
+        if let Some(last) = self.last_arrival {
+            let interval_ms = now.duration_since(last).as_secs_f64() * 1000.0;
+            let expected_ms = pkt.header.codec_id.frame_duration_ms() as f64;
+            let diff = (interval_ms - expected_ms).abs();
+            self.jitter_ms = 0.1 * diff + 0.9 * self.jitter_ms;
+        }
+        self.last_arrival = Some(now);
+    }
+
+    fn loss_percent(&self) -> f64 {
+        let total = self.packets + self.lost;
+        if total == 0 {
+            0.0
+        } else {
+            (self.lost as f64 / total as f64) * 100.0
+        }
+    }
+
+    fn duration(&self) -> Duration {
+        self.last_seen.duration_since(self.first_seen)
+    }
+
+    fn display_name(&self) -> String {
+        self.alias
+            .as_deref()
+            .map(String::from)
+            .unwrap_or_else(|| format!("Stream {}", self.stream_id))
+    }
+}
+
+// ---------------------------------------------------------------------------
+// Participant identification by sequence stream
+// ---------------------------------------------------------------------------
+
+/// Find the participant whose sequence counter is close to `seq`, or create a
+/// new one.  Each sender has an independent wrapping u16 counter, so we can
+/// distinguish streams by proximity of consecutive sequence numbers.
+fn find_or_create_participant(
+    participants: &mut Vec<ParticipantStats>,
+    seq: u16,
+    codec: CodecId,
+) -> usize {
+    for (i, p) in participants.iter().enumerate() {
+        if p.seq_initialized {
+            let delta = seq.wrapping_sub(p.last_seq);
+            if delta > 0 && delta < 50 {
+                return i;
+            }
+        }
+    }
+    // New stream detected
+    let id = participants.len();
+    participants.push(ParticipantStats::new(id, codec));
+    id
+}
+
+// ---------------------------------------------------------------------------
+// Capture writer (binary packet log for later replay)
+// ---------------------------------------------------------------------------
+
+struct CaptureWriter {
+    file: std::io::BufWriter<std::fs::File>,
+    start: Instant,
+}
+
+impl CaptureWriter {
+    fn new(path: &str, room: &str, relay: &str) -> anyhow::Result<Self> {
+        let file = std::fs::File::create(path)?;
+        let mut writer = std::io::BufWriter::new(file);
+        // Magic + version
+        writer.write_all(b"WZP\x01")?;
+        let header = serde_json::json!({
+            "room": room,
+            "relay": relay,
+            "start_time": chrono::Utc::now().to_rfc3339(),
+            "version": 1,
+        });
+        let header_bytes = serde_json::to_vec(&header)?;
+        writer.write_all(&(header_bytes.len() as u32).to_le_bytes())?;
+        writer.write_all(&header_bytes)?;
+        Ok(Self {
+            file: writer,
+            start: Instant::now(),
+        })
+    }
+
+    fn write_packet(&mut self, pkt: &MediaPacket, now: Instant) -> anyhow::Result<()> {
+        let elapsed_us = now.duration_since(self.start).as_micros() as u64;
+        self.file.write_all(&elapsed_us.to_le_bytes())?;
+        let raw = pkt.to_bytes();
+        self.file.write_all(&(raw.len() as u32).to_le_bytes())?;
+        self.file.write_all(&raw)?;
+        Ok(())
+    }
+}
+
+// ---------------------------------------------------------------------------
+// Capture reader (for replay mode)
+// ---------------------------------------------------------------------------
+
+struct CaptureReader {
+    reader: std::io::BufReader<std::fs::File>,
+    header: serde_json::Value,
+}
+
+impl CaptureReader {
+    fn open(path: &str) -> anyhow::Result<Self> {
+        use std::io::Read;
+        let file = std::fs::File::open(path)?;
+        let mut reader = std::io::BufReader::new(file);
+
+        // Read magic
+        let mut magic = [0u8; 4];
+        reader.read_exact(&mut magic)?;
+        anyhow::ensure!(&magic == b"WZP\x01", "not a WZP capture file");
+
+        // Read header
+        let mut len_buf = [0u8; 4];
+        reader.read_exact(&mut len_buf)?;
+        let header_len = u32::from_le_bytes(len_buf) as usize;
+        let mut header_bytes = vec![0u8; header_len];
+        reader.read_exact(&mut header_bytes)?;
+        let header: serde_json::Value = serde_json::from_slice(&header_bytes)?;
+
+        Ok(Self { reader, header })
+    }
+
+    fn next_packet(&mut self) -> anyhow::Result<Option<(u64, MediaPacket)>> {
+        use std::io::Read;
+        // Read timestamp
+        let mut ts_buf = [0u8; 8];
+        match self.reader.read_exact(&mut ts_buf) {
+            Ok(()) => {}
+            Err(e) if e.kind() == std::io::ErrorKind::UnexpectedEof => return Ok(None),
+            Err(e) => return Err(e.into()),
+        }
+        let timestamp_us = u64::from_le_bytes(ts_buf);
+
+        // Read packet
+        let mut len_buf = [0u8; 4];
+        self.reader.read_exact(&mut len_buf)?;
+        let pkt_len = u32::from_le_bytes(len_buf) as usize;
+        let mut pkt_bytes = vec![0u8; pkt_len];
+        self.reader.read_exact(&mut pkt_bytes)?;
+
+        let pkt = MediaPacket::from_bytes(bytes::Bytes::from(pkt_bytes))
+            .ok_or_else(|| anyhow::anyhow!("malformed packet in capture"))?;
+
+        Ok(Some((timestamp_us, pkt)))
+    }
+}
+
+// ---------------------------------------------------------------------------
+// Timeline entry (for HTML report generation)
+// ---------------------------------------------------------------------------
+
+struct TimelineEntry {
+    timestamp_us: u64,
+    stream_id: usize,
+    #[allow(dead_code)]
+    codec: CodecId,
+    #[allow(dead_code)]
+    seq: u16,
+    #[allow(dead_code)]
+    payload_len: usize,
+    loss_pct: f64,
+    jitter_ms: f64,
+}
+
+// ---------------------------------------------------------------------------
+// Replay mode (#15)
+// ---------------------------------------------------------------------------
+
+async fn run_replay(path: &str, args: &Args) -> anyhow::Result<()> {
+    let mut reader = CaptureReader::open(path)?;
+    eprintln!(
+        "Replaying: {} (room: {})",
+        path,
+        reader
+            .header
+            .get("room")
+            .and_then(|v| v.as_str())
+            .unwrap_or("?")
+    );
+
+    let mut participants: Vec<ParticipantStats> = Vec::new();
+    let mut total_packets: u64 = 0;
+    let start = Instant::now();
+    let mut timeline: Vec<TimelineEntry> = Vec::new();
+
+    // Decrypt session from --key (optional)
+    let mut decrypt_session: Option<wzp_crypto::ChaChaSession> = args.key.as_ref().and_then(|hex| {
+        if hex.len() != 64 { return None; }
+        let mut key = [0u8; 32];
+        for (i, chunk) in hex.as_bytes().chunks(2).enumerate() {
+            let s = std::str::from_utf8(chunk).unwrap_or("00");
+            key[i] = u8::from_str_radix(s, 16).unwrap_or(0);
+        }
+        Some(wzp_crypto::ChaChaSession::new(key))
+    });
+    let mut decrypt_ok: u64 = 0;
+    let mut decrypt_fail: u64 = 0;
+
+    while let Some((ts_us, pkt)) = reader.next_packet()? {
+        let now = Instant::now();
+        let idx = find_or_create_participant(&mut participants, pkt.header.seq, pkt.header.codec_id);
+        participants[idx].ingest(&pkt, now);
+        total_packets += 1;
+
+        // Attempt decryption if key provided
+        if let Some(ref mut session) = decrypt_session {
+            use wzp_proto::CryptoSession;
+            let header_bytes = pkt.header.to_bytes();
+            let mut plaintext = Vec::new();
+            match session.decrypt(&header_bytes, &pkt.payload, &mut plaintext) {
+                Ok(()) => {
+                    decrypt_ok += 1;
+                    if decrypt_ok <= 5 || decrypt_ok % 100 == 0 {
+                        eprintln!(
+                            "  decrypt ok: seq={} codec={:?} payload={}B → plaintext={}B",
+                            pkt.header.seq, pkt.header.codec_id,
+                            pkt.payload.len(), plaintext.len()
+                        );
+                    }
+                }
+                Err(_) => {
+                    decrypt_fail += 1;
+                    if decrypt_fail <= 3 {
+                        eprintln!(
+                            "  decrypt FAIL: seq={} (key mismatch, wrong direction, or rekey boundary)",
+                            pkt.header.seq
+                        );
+                    }
+                }
+            }
+        }
+
+        // Record for HTML timeline
+        timeline.push(TimelineEntry {
+            timestamp_us: ts_us,
+            stream_id: idx,
+            codec: pkt.header.codec_id,
+            seq: pkt.header.seq,
+            payload_len: pkt.payload.len(),
+            loss_pct: participants[idx].loss_percent(),
+            jitter_ms: participants[idx].jitter_ms,
+        });
+    }
+
+    if decrypt_session.is_some() {
+        eprintln!(
+            "Decrypt stats: {} ok, {} failed (total {})",
+            decrypt_ok, decrypt_fail, total_packets
+        );
+    }
+
+    print_summary(&participants, total_packets, start.elapsed());
+
+    // Generate HTML if requested
+    if let Some(html_path) = &args.html {
+        generate_html_report(html_path, &participants, &timeline, total_packets, &reader.header)?;
+        eprintln!("HTML report: {}", html_path);
+    }
+
+    Ok(())
+}
+
+// ---------------------------------------------------------------------------
+// HTML report generation (#16)
+// ---------------------------------------------------------------------------
+
+fn generate_html_report(
+    path: &str,
+    participants: &[ParticipantStats],
+    timeline: &[TimelineEntry],
+    total_packets: u64,
+    capture_header: &serde_json::Value,
+) -> anyhow::Result<()> {
+    use std::io::Write as _;
+    let mut f = std::fs::File::create(path)?;
+
+    let room = capture_header
+        .get("room")
+        .and_then(|v| v.as_str())
+        .unwrap_or("unknown");
+    let start_time = capture_header
+        .get("start_time")
+        .and_then(|v| v.as_str())
+        .unwrap_or("?");
+
+    // Build per-stream loss/jitter timeline data for Chart.js
+    // Sample every 1 second (group timeline entries by second)
+    let max_ts = timeline.last().map(|e| e.timestamp_us).unwrap_or(0);
+    let duration_secs = (max_ts / 1_000_000) + 1;
+
+    let mut loss_data: std::collections::HashMap<usize, Vec<f64>> =
+        std::collections::HashMap::new();
+    let mut jitter_data: std::collections::HashMap<usize, Vec<f64>> =
+        std::collections::HashMap::new();
+
+    for stream_id in 0..participants.len() {
+        loss_data.insert(stream_id, vec![0.0; duration_secs as usize]);
+        jitter_data.insert(stream_id, vec![0.0; duration_secs as usize]);
+    }
+
+    for entry in timeline {
+        let sec = (entry.timestamp_us / 1_000_000) as usize;
+        if sec < duration_secs as usize {
+            if let Some(losses) = loss_data.get_mut(&entry.stream_id) {
+                losses[sec] = entry.loss_pct;
+            }
+            if let Some(jitters) = jitter_data.get_mut(&entry.stream_id) {
+                jitters[sec] = entry.jitter_ms;
+            }
+        }
+    }
+
+    let colors = [
+        "#e74c3c", "#3498db", "#2ecc71", "#f39c12", "#9b59b6", "#1abc9c",
+    ];
+
+    // Build dataset JSON for charts
+    let mut loss_datasets = String::new();
+    let mut jitter_datasets = String::new();
+    for (i, p) in participants.iter().enumerate() {
+        let name = p.display_name();
+        let color = colors[i % colors.len()];
+        let loss_vals = loss_data
+            .get(&i)
+            .map(|v| format!("{:?}", v))
+            .unwrap_or_default();
+        let jitter_vals = jitter_data
+            .get(&i)
+            .map(|v| format!("{:?}", v))
+            .unwrap_or_default();
+
+        loss_datasets.push_str(&format!(
+            "{{ label: '{}', data: {}, borderColor: '{}', fill: false }},\n",
+            name, loss_vals, color
+        ));
+        jitter_datasets.push_str(&format!(
+            "{{ label: '{}', data: {}, borderColor: '{}', fill: false }},\n",
+            name, jitter_vals, color
+        ));
+    }
+
+    let labels: Vec<String> = (0..duration_secs).map(|s| format!("{}s", s)).collect();
+    let labels_json = format!("{:?}", labels);
+
+    // Summary table rows
+    let mut summary_rows = String::new();
+    for p in participants {
+        summary_rows.push_str(&format!(
+            "<tr><td>{}</td><td>{:?}</td><td>{}</td><td>{:.1}%</td><td>{:.0}ms</td><td>{}</td></tr>\n",
+            p.display_name(),
+            p.codec,
+            p.packets,
+            p.loss_percent(),
+            p.jitter_ms,
+            p.codec_switches
+        ));
+    }
+
+    write!(
+        f,
+        r#"<!DOCTYPE html>
+<html><head>
+<meta charset="utf-8">
+<title>WZP Call Report — {room}</title>
+<script src="https://cdn.jsdelivr.net/npm/chart.js@4"></script>
+<style>
+  body {{ font-family: -apple-system, sans-serif; max-width: 1200px; margin: 0 auto; padding: 20px; background: #1a1a2e; color: #e0e0e0; }}
+  h1,h2 {{ color: #4a9eff; }}
+  table {{ border-collapse: collapse; width: 100%; margin: 20px 0; }}
+  th,td {{ border: 1px solid #333; padding: 8px 12px; text-align: left; }}
+  th {{ background: #16213e; }}
+  tr:nth-child(even) {{ background: #1a1a3e; }}
+  .chart-container {{ background: #16213e; border-radius: 8px; padding: 16px; margin: 20px 0; }}
+  canvas {{ max-height: 300px; }}
+  .meta {{ color: #888; font-size: 0.9em; }}
+</style>
+</head><body>
+<h1>WZP Call Quality Report</h1>
+<p class="meta">Room: <b>{room}</b> | Start: {start_time} | Packets: {total_packets} | Duration: {duration_secs}s</p>
+
+<h2>Participant Summary</h2>
+<table>
+<tr><th>Name</th><th>Codec</th><th>Packets</th><th>Loss</th><th>Jitter</th><th>Codec Switches</th></tr>
+{summary_rows}
+</table>
+
+<h2>Packet Loss Over Time</h2>
+<div class="chart-container"><canvas id="lossChart"></canvas></div>
+
+<h2>Jitter Over Time</h2>
+<div class="chart-container"><canvas id="jitterChart"></canvas></div>
+
+<script>
+const labels = {labels_json};
+new Chart(document.getElementById('lossChart'), {{
+  type: 'line',
+  data: {{ labels, datasets: [{loss_datasets}] }},
+  options: {{ responsive: true, scales: {{ y: {{ beginAtZero: true, title: {{ display: true, text: 'Loss %' }} }} }} }}
+}});
+new Chart(document.getElementById('jitterChart'), {{
+  type: 'line',
+  data: {{ labels, datasets: [{jitter_datasets}] }},
+  options: {{ responsive: true, scales: {{ y: {{ beginAtZero: true, title: {{ display: true, text: 'Jitter (ms)' }} }} }} }}
+}});
+</script>
+</body></html>"#
+    )?;
+
+    Ok(())
+}
+
+// ---------------------------------------------------------------------------
+// No-TUI mode (print stats to stdout periodically)
+// ---------------------------------------------------------------------------
+
+async fn run_no_tui(
+    transport: &wzp_transport::QuinnTransport,
+    participants: &mut Vec<ParticipantStats>,
+    total_packets: &mut u64,
+    deadline: Option<Instant>,
+    mut capture_writer: Option<&mut CaptureWriter>,
+) -> anyhow::Result<()> {
+    let mut print_timer = Instant::now();
+    loop {
+        if let Some(dl) = deadline {
+            if Instant::now() > dl {
+                break;
+            }
+        }
+        match tokio::time::timeout(Duration::from_millis(100), transport.recv_media()).await {
+            Ok(Ok(Some(pkt))) => {
+                let now = Instant::now();
+                let idx =
+                    find_or_create_participant(participants, pkt.header.seq, pkt.header.codec_id);
+                participants[idx].ingest(&pkt, now);
+                *total_packets += 1;
+                if let Some(ref mut w) = capture_writer {
+                    w.write_packet(&pkt, now)?;
+                }
+            }
+            Ok(Ok(None)) => break,   // connection closed
+            Ok(Err(e)) => {
+                tracing::warn!("recv error: {e}");
+                break;
+            }
+            Err(_) => {}              // timeout, loop again
+        }
+        if print_timer.elapsed() >= Duration::from_secs(2) {
+            print_stats(participants, *total_packets);
+            print_timer = Instant::now();
+        }
+    }
+    Ok(())
+}
+
+fn print_stats(participants: &[ParticipantStats], total: u64) {
+    eprintln!("--- {} participants | {} total packets ---", participants.len(), total);
+    for p in participants {
+        eprintln!(
+            "  {}: {} pkts, {:.1}% loss, {:.0}ms jitter, {:?}, {:.0}s",
+            p.display_name(),
+            p.packets,
+            p.loss_percent(),
+            p.jitter_ms,
+            p.codec,
+            p.duration().as_secs_f64(),
+        );
+    }
+}
+
+// ---------------------------------------------------------------------------
+// TUI mode (ratatui + crossterm)
+// ---------------------------------------------------------------------------
+
+async fn run_tui(
+    transport: &wzp_transport::QuinnTransport,
+    participants: &mut Vec<ParticipantStats>,
+    total_packets: &mut u64,
+    start_time: Instant,
+    deadline: Option<Instant>,
+    mut capture_writer: Option<&mut CaptureWriter>,
+) -> anyhow::Result<()> {
+    crossterm::terminal::enable_raw_mode()?;
+    let mut stdout = std::io::stdout();
+    crossterm::execute!(stdout, crossterm::terminal::EnterAlternateScreen)?;
+    let backend = ratatui::backend::CrosstermBackend::new(stdout);
+    let mut terminal = ratatui::Terminal::new(backend)?;
+
+    let mut redraw_timer = Instant::now();
+
+    let result: anyhow::Result<()> = async {
+        loop {
+            // Check for quit key (q or Ctrl+C)
+            if crossterm::event::poll(Duration::from_millis(0))? {
+                if let crossterm::event::Event::Key(key) = crossterm::event::read()? {
+                    use crossterm::event::{KeyCode, KeyModifiers};
+                    if key.code == KeyCode::Char('q')
+                        || (key.code == KeyCode::Char('c')
+                            && key.modifiers.contains(KeyModifiers::CONTROL))
+                    {
+                        break;
+                    }
+                }
+            }
+
+            if let Some(dl) = deadline {
+                if Instant::now() > dl {
+                    break;
+                }
+            }
+
+            // Receive packets (non-blocking with short timeout)
+            match tokio::time::timeout(Duration::from_millis(20), transport.recv_media()).await {
+                Ok(Ok(Some(pkt))) => {
+                    let now = Instant::now();
+                    let idx = find_or_create_participant(
+                        participants,
+                        pkt.header.seq,
+                        pkt.header.codec_id,
+                    );
+                    participants[idx].ingest(&pkt, now);
+                    *total_packets += 1;
+                    if let Some(ref mut w) = capture_writer {
+                        w.write_packet(&pkt, now)?;
+                    }
+                }
+                Ok(Ok(None)) => break,
+                Ok(Err(e)) => {
+                    tracing::warn!("recv error: {e}");
+                    break;
+                }
+                Err(_) => {}
+            }
+
+            // Redraw TUI at ~10 FPS
+            if redraw_timer.elapsed() >= Duration::from_millis(100) {
+                terminal.draw(|f| draw_ui(f, participants, *total_packets, start_time))?;
+                redraw_timer = Instant::now();
+            }
+        }
+        Ok(())
+    }
+    .await;
+
+    // Always restore terminal, even on error
+    crossterm::terminal::disable_raw_mode()?;
+    crossterm::execute!(
+        std::io::stdout(),
+        crossterm::terminal::LeaveAlternateScreen
+    )?;
+
+    result
+}
+
+fn draw_ui(
+    f: &mut ratatui::Frame,
+    participants: &[ParticipantStats],
+    total_packets: u64,
+    start_time: Instant,
+) {
+    use ratatui::layout::{Constraint, Direction, Layout};
+    use ratatui::style::{Color, Modifier, Style};
+    use ratatui::widgets::{Block, Borders, Paragraph, Row, Table};
+
+    let elapsed = start_time.elapsed();
+    let elapsed_str = format!(
+        "{:02}:{:02}:{:02}",
+        elapsed.as_secs() / 3600,
+        (elapsed.as_secs() % 3600) / 60,
+        elapsed.as_secs() % 60
+    );
+
+    let chunks = Layout::default()
+        .direction(Direction::Vertical)
+        .constraints([
+            Constraint::Length(3), // header
+            Constraint::Min(5),   // participant table
+            Constraint::Length(3), // footer
+        ])
+        .split(f.area());
+
+    // Header
+    let header = Paragraph::new(format!(
+        " WZP Analyzer | {} participants | {} packets | {}",
+        participants.len(),
+        total_packets,
+        elapsed_str
+    ))
+    .block(Block::default().borders(Borders::ALL).title(" Protocol Analyzer "));
+    f.render_widget(header, chunks[0]);
+
+    // Participant table
+    let header_row = Row::new(vec![
+        "#", "Name", "Codec", "Packets", "Loss%", "Jitter", "Switches", "Duration",
+    ])
+    .style(Style::default().add_modifier(Modifier::BOLD));
+
+    let rows: Vec<Row> = participants
+        .iter()
+        .map(|p| {
+            let loss_color = if p.loss_percent() > 5.0 {
+                Color::Red
+            } else if p.loss_percent() > 1.0 {
+                Color::Yellow
+            } else {
+                Color::Green
+            };
+
+            Row::new(vec![
+                format!("{}", p.stream_id),
+                p.display_name(),
+                format!("{:?}", p.codec),
+                format!("{}", p.packets),
+                format!("{:.1}%", p.loss_percent()),
+                format!("{:.0}ms", p.jitter_ms),
+                format!("{}", p.codec_switches),
+                format!("{:.0}s", p.duration().as_secs_f64()),
+            ])
+            .style(Style::default().fg(loss_color))
+        })
+        .collect();
+
+    let widths = [
+        Constraint::Length(3),  // #
+        Constraint::Length(20), // Name
+        Constraint::Length(12), // Codec
+        Constraint::Length(10), // Packets
+        Constraint::Length(8),  // Loss%
+        Constraint::Length(10), // Jitter
+        Constraint::Length(10), // Switches
+        Constraint::Length(10), // Duration
+    ];
+
+    let table = Table::new(rows, widths)
+        .header(header_row)
+        .block(Block::default().borders(Borders::ALL).title(" Participants "));
+    f.render_widget(table, chunks[1]);
+
+    // Footer
+    let footer =
+        Paragraph::new(" Press 'q' to quit ").block(Block::default().borders(Borders::ALL));
+    f.render_widget(footer, chunks[2]);
+}
+
+// ---------------------------------------------------------------------------
+// Summary (printed on exit)
+// ---------------------------------------------------------------------------
+
+fn print_summary(participants: &[ParticipantStats], total: u64, elapsed: Duration) {
+    eprintln!("\n=== Session Summary ===");
+    eprintln!(
+        "Duration: {:.1}s | Total packets: {} | Participants: {}",
+        elapsed.as_secs_f64(),
+        total,
+        participants.len()
+    );
+    for p in participants {
+        eprintln!(
+            "  {}: {} pkts, {:.1}% loss, {:.0}ms jitter, {:?}, {} codec switches",
+            p.display_name(),
+            p.packets,
+            p.loss_percent(),
+            p.jitter_ms,
+            p.codec,
+            p.codec_switches,
+        );
+    }
+}
+
+// ---------------------------------------------------------------------------
+// main
+// ---------------------------------------------------------------------------
+
+#[tokio::main]
+async fn main() -> anyhow::Result<()> {
+    let args = Args::parse();
+
+    // Only init tracing subscriber in no-tui mode (it would corrupt the TUI otherwise)
+    if args.no_tui || args.replay.is_some() {
+        tracing_subscriber::fmt().init();
+    }
+
+    let _crypto_session: Option<std::sync::Mutex<wzp_crypto::ChaChaSession>> =
+        if let Some(ref key_hex) = args.key {
+            if key_hex.len() != 64 {
+                eprintln!("Error: --key must be 64 hex characters (32 bytes). Got {} chars.", key_hex.len());
+                std::process::exit(1);
+            }
+            let mut key_bytes = [0u8; 32];
+            for (i, chunk) in key_hex.as_bytes().chunks(2).enumerate() {
+                let hex_str = std::str::from_utf8(chunk).unwrap_or("00");
+                key_bytes[i] = u8::from_str_radix(hex_str, 16).unwrap_or(0);
+            }
+            eprintln!("Encrypted payload decoding enabled (key loaded).");
+            Some(std::sync::Mutex::new(
+                wzp_crypto::ChaChaSession::new(key_bytes),
+            ))
+        } else {
+            None
+        };
+
+    // Replay mode: offline analysis of a .wzp capture file
+    if let Some(ref replay_path) = args.replay {
+        return run_replay(replay_path, &args).await;
+    }
+
+    // Live mode requires relay and room
+    let relay = args
+        .relay
+        .as_deref()
+        .ok_or_else(|| anyhow::anyhow!("relay address required for live mode (use --replay for offline)"))?;
+    let room = args
+        .room
+        .as_deref()
+        .ok_or_else(|| anyhow::anyhow!("--room required for live mode (use --replay for offline)"))?;
+
+    // TLS crypto provider
+    let _ = rustls::crypto::ring::default_provider().install_default();
+
+    // Identity seed
+    let seed = match &args.seed {
+        Some(hex) => {
+            let s = wzp_crypto::Seed::from_hex(hex).map_err(|e| anyhow::anyhow!(e))?;
+            info!(fingerprint = %s.derive_identity().public_identity().fingerprint, "identity from --seed");
+            s
+        }
+        None => {
+            let s = wzp_crypto::Seed::generate();
+            info!(fingerprint = %s.derive_identity().public_identity().fingerprint, "generated ephemeral identity");
+            s
+        }
+    };
+
+    // Connect to relay
+    let relay_addr: std::net::SocketAddr = relay.parse()?;
+    let bind_addr: std::net::SocketAddr = if relay_addr.is_ipv6() {
+        "[::]:0".parse()?
+    } else {
+        "0.0.0.0:0".parse()?
+    };
+    let endpoint = wzp_transport::create_endpoint(bind_addr, None)?;
+    let client_config = wzp_transport::client_config();
+    let conn = wzp_transport::connect(&endpoint, relay_addr, room, client_config).await?;
+    let transport = Arc::new(wzp_transport::QuinnTransport::new(conn));
+
+    // Crypto handshake
+    let _crypto_session =
+        wzp_client::handshake::perform_handshake(&*transport, &seed.0, Some("analyzer")).await?;
+
+    // Auth if token provided
+    if let Some(ref token) = args.token {
+        let auth = wzp_proto::SignalMessage::AuthToken {
+            token: token.clone(),
+        };
+        transport.send_signal(&auth).await?;
+    }
+
+    // Capture file (optional)
+    let mut capture_writer = args
+        .capture
+        .as_ref()
+        .map(|path| CaptureWriter::new(path, room, relay))
+        .transpose()?;
+
+    // Duration timeout
+    let deadline = args
+        .duration
+        .map(|s| Instant::now() + Duration::from_secs(s));
+
+    // State
+    let mut participants: Vec<ParticipantStats> = Vec::new();
+    let mut total_packets: u64 = 0;
+    let start_time = Instant::now();
+
+    if args.no_tui {
+        run_no_tui(
+            &transport,
+            &mut participants,
+            &mut total_packets,
+            deadline,
+            capture_writer.as_mut(),
+        )
+        .await?;
+    } else {
+        run_tui(
+            &transport,
+            &mut participants,
+            &mut total_packets,
+            start_time,
+            deadline,
+            capture_writer.as_mut(),
+        )
+        .await?;
+    }
+
+    // Print summary
+    print_summary(&participants, total_packets, start_time.elapsed());
+
+    // Clean close
+    transport.close().await?;
+
+    Ok(())
+}

crates/wzp-client/src/audio_io.rs

@@ -3,12 +3,10 @@
 //! Both structs use 48 kHz, mono, i16 format to match the WarzonePhone codec
 //! pipeline. Frames are 960 samples (20 ms at 48 kHz).
 //!
-//! The cpal `Stream` type is not `Send`, so each struct spawns a dedicated OS
-//! thread that owns the stream. The public API exposes only `Send + Sync`
-//! channel handles.
+//! Audio callbacks are **lock-free**: they read/write directly to an `AudioRing`
+//! (atomic SPSC ring buffer). No Mutex, no channel, no allocation on the hot path.
 
 use std::sync::atomic::{AtomicBool, Ordering};
-use std::sync::mpsc;
 use std::sync::Arc;
 
 use anyhow::{anyhow, Context};
@@ -16,6 +14,8 @@ use cpal::traits::{DeviceTrait, HostTrait, StreamTrait};
 use cpal::{SampleFormat, SampleRate, StreamConfig};
 use tracing::{info, warn};
 
+use crate::audio_ring::AudioRing;
+
 /// Number of samples per 20 ms frame at 48 kHz mono.
 pub const FRAME_SAMPLES: usize = 960;
 
@@ -23,22 +23,24 @@ pub const FRAME_SAMPLES: usize = 960;
 // AudioCapture
 // ---------------------------------------------------------------------------
 
-/// Captures microphone input and yields 960-sample PCM frames.
+/// Captures microphone input via CPAL and writes PCM into a lock-free ring buffer.
 ///
 /// The cpal stream lives on a dedicated OS thread; this handle is `Send + Sync`.
 pub struct AudioCapture {
-    rx: mpsc::Receiver<Vec<i16>>,
+    ring: Arc<AudioRing>,
     running: Arc<AtomicBool>,
 }
 
 impl AudioCapture {
     /// Create and start capturing from the default input device at 48 kHz mono.
     pub fn start() -> Result<Self, anyhow::Error> {
-        let (tx, rx) = mpsc::sync_channel::<Vec<i16>>(64);
+        let ring = Arc::new(AudioRing::new());
         let running = Arc::new(AtomicBool::new(true));
-        let running_clone = running.clone();
 
-        let (init_tx, init_rx) = mpsc::sync_channel::<Result<(), String>>(1);
+        let (init_tx, init_rx) = std::sync::mpsc::sync_channel::<Result<(), String>>(1);
+
+        let ring_cb = ring.clone();
+        let running_clone = running.clone();
 
         std::thread::Builder::new()
             .name("wzp-audio-capture".into())
@@ -59,53 +61,51 @@ impl AudioCapture {
 
                     let use_f32 = !supports_i16_input(&device)?;
 
-                    let buf = Arc::new(std::sync::Mutex::new(
-                        Vec::<i16>::with_capacity(FRAME_SAMPLES),
-                    ));
                     let err_cb = |e: cpal::StreamError| {
                         warn!("input stream error: {e}");
                     };
 
+                    let logged_cb_size = Arc::new(AtomicBool::new(false));
+
                     let stream = if use_f32 {
-                        let buf = buf.clone();
-                        let tx = tx.clone();
+                        let ring = ring_cb.clone();
                         let running = running_clone.clone();
+                        let logged = logged_cb_size.clone();
                         device.build_input_stream(
                             &config,
                             move |data: &[f32], _: &cpal::InputCallbackInfo| {
                                 if !running.load(Ordering::Relaxed) {
                                     return;
                                 }
-                                let mut lock = buf.lock().unwrap();
-                                for &s in data {
-                                    lock.push(f32_to_i16(s));
-                                    if lock.len() == FRAME_SAMPLES {
-                                        let frame = lock.drain(..).collect();
-                                        let _ = tx.try_send(frame);
+                                if !logged.swap(true, Ordering::Relaxed) {
+                                    eprintln!("[audio] capture callback: {} f32 samples", data.len());
+                                }
+                                let mut tmp = [0i16; FRAME_SAMPLES];
+                                for chunk in data.chunks(FRAME_SAMPLES) {
+                                    let n = chunk.len();
+                                    for i in 0..n {
+                                        tmp[i] = f32_to_i16(chunk[i]);
                                     }
+                                    ring.write(&tmp[..n]);
                                 }
                             },
                             err_cb,
                             None,
                         )?
                     } else {
-                        let buf = buf.clone();
-                        let tx = tx.clone();
+                        let ring = ring_cb.clone();
                         let running = running_clone.clone();
+                        let logged = logged_cb_size.clone();
                         device.build_input_stream(
                             &config,
                             move |data: &[i16], _: &cpal::InputCallbackInfo| {
                                 if !running.load(Ordering::Relaxed) {
                                     return;
                                 }
-                                let mut lock = buf.lock().unwrap();
-                                for &s in data {
-                                    lock.push(s);
-                                    if lock.len() == FRAME_SAMPLES {
-                                        let frame = lock.drain(..).collect();
-                                        let _ = tx.try_send(frame);
-                                    }
+                                if !logged.swap(true, Ordering::Relaxed) {
+                                    eprintln!("[audio] capture callback: {} i16 samples", data.len());
                                 }
+                                ring.write(data);
                             },
                             err_cb,
                             None,
@@ -114,7 +114,6 @@ impl AudioCapture {
 
                     stream.play().context("failed to start input stream")?;
 
-                    // Signal success to the caller before parking.
                     let _ = init_tx.send(Ok(()));
 
                     // Keep stream alive until stopped.
@@ -135,15 +134,12 @@ impl AudioCapture {
             .map_err(|_| anyhow!("capture thread exited before signaling"))?
             .map_err(|e| anyhow!("{e}"))?;
 
-        Ok(Self { rx, running })
+        Ok(Self { ring, running })
     }
 
-    /// Read the next frame of 960 PCM samples (blocking until available).
-    ///
-    /// Returns `None` when the stream has been stopped or the channel is
-    /// disconnected.
-    pub fn read_frame(&self) -> Option<Vec<i16>> {
-        self.rx.recv().ok()
+    /// Get a reference to the capture ring buffer for direct polling.
+    pub fn ring(&self) -> &Arc<AudioRing> {
+        &self.ring
     }
 
     /// Stop capturing.
@@ -152,26 +148,34 @@ impl AudioCapture {
     }
 }
 
+impl Drop for AudioCapture {
+    fn drop(&mut self) {
+        self.stop();
+    }
+}
+
 // ---------------------------------------------------------------------------
 // AudioPlayback
 // ---------------------------------------------------------------------------
 
-/// Plays PCM frames through the default output device at 48 kHz mono.
+/// Plays PCM through the default output device, reading from a lock-free ring buffer.
 ///
 /// The cpal stream lives on a dedicated OS thread; this handle is `Send + Sync`.
 pub struct AudioPlayback {
-    tx: mpsc::SyncSender<Vec<i16>>,
+    ring: Arc<AudioRing>,
     running: Arc<AtomicBool>,
 }
 
 impl AudioPlayback {
     /// Create and start playback on the default output device at 48 kHz mono.
     pub fn start() -> Result<Self, anyhow::Error> {
-        let (tx, rx) = mpsc::sync_channel::<Vec<i16>>(64);
+        let ring = Arc::new(AudioRing::new());
         let running = Arc::new(AtomicBool::new(true));
-        let running_clone = running.clone();
 
-        let (init_tx, init_rx) = mpsc::sync_channel::<Result<(), String>>(1);
+        let (init_tx, init_rx) = std::sync::mpsc::sync_channel::<Result<(), String>>(1);
+
+        let ring_cb = ring.clone();
+        let running_clone = running.clone();
 
         std::thread::Builder::new()
             .name("wzp-audio-playback".into())
@@ -192,62 +196,40 @@ impl AudioPlayback {
 
                     let use_f32 = !supports_i16_output(&device)?;
 
-                    // Shared ring of samples the cpal callback drains from.
-                    let ring = Arc::new(std::sync::Mutex::new(
-                        std::collections::VecDeque::<i16>::with_capacity(FRAME_SAMPLES * 8),
-                    ));
-
-                    // Background drainer: moves frames from the mpsc channel into the ring.
-                    {
-                        let ring = ring.clone();
-                        let running = running_clone.clone();
-                        std::thread::Builder::new()
-                            .name("wzp-playback-drain".into())
-                            .spawn(move || {
-                                while running.load(Ordering::Relaxed) {
-                                    match rx.recv_timeout(std::time::Duration::from_millis(100)) {
-                                        Ok(frame) => {
-                                            let mut lock = ring.lock().unwrap();
-                                            lock.extend(frame);
-                                            while lock.len() > FRAME_SAMPLES * 16 {
-                                                lock.pop_front();
-                                            }
-                                        }
-                                        Err(mpsc::RecvTimeoutError::Timeout) => {}
-                                        Err(mpsc::RecvTimeoutError::Disconnected) => break,
-                                    }
-                                }
-                            })?;
-                    }
-
                     let err_cb = |e: cpal::StreamError| {
                         warn!("output stream error: {e}");
                     };
 
                     let stream = if use_f32 {
-                        let ring = ring.clone();
+                        let ring = ring_cb.clone();
                         device.build_output_stream(
                             &config,
                             move |data: &mut [f32], _: &cpal::OutputCallbackInfo| {
-                                let mut lock = ring.lock().unwrap();
-                                for sample in data.iter_mut() {
-                                    *sample = match lock.pop_front() {
-                                        Some(s) => i16_to_f32(s),
-                                        None => 0.0,
-                                    };
+                                let mut tmp = [0i16; FRAME_SAMPLES];
+                                for chunk in data.chunks_mut(FRAME_SAMPLES) {
+                                    let n = chunk.len();
+                                    let read = ring.read(&mut tmp[..n]);
+                                    for i in 0..read {
+                                        chunk[i] = i16_to_f32(tmp[i]);
+                                    }
+                                    // Fill remainder with silence if ring underran
+                                    for i in read..n {
+                                        chunk[i] = 0.0;
+                                    }
                                 }
                             },
                             err_cb,
                             None,
                         )?
                     } else {
-                        let ring = ring.clone();
+                        let ring = ring_cb.clone();
                         device.build_output_stream(
                             &config,
                             move |data: &mut [i16], _: &cpal::OutputCallbackInfo| {
-                                let mut lock = ring.lock().unwrap();
-                                for sample in data.iter_mut() {
-                                    *sample = lock.pop_front().unwrap_or(0);
+                                let read = ring.read(data);
+                                // Fill remainder with silence if ring underran
+                                for sample in &mut data[read..] {
+                                    *sample = 0;
                                 }
                             },
                             err_cb,
@@ -257,7 +239,6 @@ impl AudioPlayback {
 
                     stream.play().context("failed to start output stream")?;
 
-                    // Signal success to the caller before parking.
                     let _ = init_tx.send(Ok(()));
 
                     // Keep stream alive until stopped.
@@ -278,12 +259,12 @@ impl AudioPlayback {
             .map_err(|_| anyhow!("playback thread exited before signaling"))?
             .map_err(|e| anyhow!("{e}"))?;
 
-        Ok(Self { tx, running })
+        Ok(Self { ring, running })
     }
 
-    /// Write a frame of PCM samples for playback.
-    pub fn write_frame(&self, pcm: &[i16]) {
-        let _ = self.tx.try_send(pcm.to_vec());
+    /// Get a reference to the playout ring buffer for direct writing.
+    pub fn ring(&self) -> &Arc<AudioRing> {
+        &self.ring
     }
 
     /// Stop playback.
@@ -292,11 +273,16 @@ impl AudioPlayback {
     }
 }
 
+impl Drop for AudioPlayback {
+    fn drop(&mut self) {
+        self.stop();
+    }
+}
+
 // ---------------------------------------------------------------------------
 // Helpers
 // ---------------------------------------------------------------------------
 
-/// Check if the input device supports i16 at 48 kHz mono.
 fn supports_i16_input(device: &cpal::Device) -> Result<bool, anyhow::Error> {
     let supported = device
         .supported_input_configs()
@@ -313,7 +299,6 @@ fn supports_i16_input(device: &cpal::Device) -> Result<bool, anyhow::Error> {
     Ok(false)
 }
 
-/// Check if the output device supports i16 at 48 kHz mono.
 fn supports_i16_output(device: &cpal::Device) -> Result<bool, anyhow::Error> {
     let supported = device
         .supported_output_configs()

crates/wzp-client/src/audio_linux_aec.rs

@@ -0,0 +1,537 @@
+//! Linux AEC backend: CPAL capture + playback wired through the WebRTC Audio
+//! Processing Module (AEC3 + noise suppression + high-pass filter).
+//!
+//! This is the same algorithm used by Chrome WebRTC, Zoom, Teams, Jitsi, and
+//! any other "serious" Linux VoIP app. It runs in-process — no dependency on
+//! PulseAudio's module-echo-cancel or PipeWire's filter-chain, so it works
+//! identically on ALSA / PulseAudio / PipeWire systems.
+//!
+//! ## Architecture
+//!
+//! A single module-level `Arc<Mutex<Processor>>` is shared between the
+//! capture and playback paths. On each 20 ms frame (960 samples @ 48 kHz
+//! mono):
+//!
+//! - **Playback path**: `LinuxAecPlayback::start` spawns the usual CPAL
+//!   output thread, but wraps each chunk in a call to
+//!   `Processor::process_render_frame` **before** handing it to CPAL. That
+//!   gives APM an authoritative reference of exactly what's going out to
+//!   the speakers (same approach Zoom/Teams/Jitsi use). The AEC then knows
+//!   what to cancel when it sees echo in the capture stream.
+//!
+//! - **Capture path**: `LinuxAecCapture::start` spawns the usual CPAL
+//!   input thread, and runs `Processor::process_capture_frame` on each
+//!   incoming mic chunk **in place** before pushing it into the ring
+//!   buffer. The AEC subtracts the echo using the render reference it
+//!   saw on the playback side.
+//!
+//! APM is strict about frame size: it requires exactly 10 ms = 480 samples
+//! per call at 48 kHz. Our pipeline uses 20 ms = 960 samples, so each 20 ms
+//! frame is split into two 480-sample halves, APM is called twice, and the
+//! halves are stitched back together.
+//!
+//! APM only accepts f32 samples in `[-1.0, 1.0]`, so we convert i16 → f32
+//! before the call and f32 → i16 after (with clamping on the return path).
+//!
+//! ## Stream delay
+//!
+//! AEC needs to know roughly how long it takes between a sample being passed
+//! to `process_render_frame` and its echo showing up at `process_capture_frame`
+//! — i.e. the round trip through CPAL playback → speaker → air → microphone
+//! → CPAL capture. AEC3's internal estimator tracks this within a window
+//! around whatever hint we give it. We hardcode 60 ms as a reasonable
+//! starting point for typical Linux audio stacks; the delay estimator does
+//! the fine-tuning automatically.
+//!
+//! ## Thread safety
+//!
+//! The 0.3.x line of `webrtc-audio-processing` takes `&mut self` on both
+//! `process_capture_frame` and `process_render_frame`, so the `Processor`
+//! needs a `Mutex` around it for cross-thread sharing. The capture and
+//! playback threads each acquire the lock briefly (sub-millisecond per
+//! 10 ms frame) so contention is minimal at our frame rates.
+
+use std::sync::atomic::{AtomicBool, Ordering};
+use std::sync::{Arc, Mutex, OnceLock};
+
+use anyhow::{anyhow, Context};
+use cpal::traits::{DeviceTrait, HostTrait, StreamTrait};
+use cpal::{SampleFormat, SampleRate, StreamConfig};
+use tracing::{info, warn};
+use webrtc_audio_processing::{
+    Config, EchoCancellation, EchoCancellationSuppressionLevel, InitializationConfig,
+    NoiseSuppression, NoiseSuppressionLevel, Processor, NUM_SAMPLES_PER_FRAME,
+};
+
+use crate::audio_ring::AudioRing;
+
+/// 20 ms at 48 kHz, mono — matches the rest of the pipeline and the codec.
+pub const FRAME_SAMPLES: usize = 960;
+/// APM requires strict 10 ms frames at 48 kHz = 480 samples per call.
+/// Imported from the webrtc-audio-processing crate so we can't drift out
+/// of sync with whatever sample rate / frame length the C++ lib is using.
+const APM_FRAME_SAMPLES: usize = NUM_SAMPLES_PER_FRAME as usize;
+const APM_NUM_CHANNELS: usize = 1;
+/// Round-trip delay hint passed to APM; the estimator refines from here.
+/// 60 ms is a reasonable default for CPAL on ALSA / PulseAudio / PipeWire.
+#[allow(dead_code)]
+const STREAM_DELAY_MS: i32 = 60;
+
+// ---------------------------------------------------------------------------
+// Shared APM instance
+// ---------------------------------------------------------------------------
+
+/// Module-level lazily-initialized APM. Shared between capture and playback
+/// so they operate on the same echo-cancellation state — the render frames
+/// pushed by playback are what the capture path subtracts from the mic input.
+/// Wrapped in a Mutex because the 0.3.x Processor takes `&mut self` on both
+/// process_capture_frame and process_render_frame.
+static PROCESSOR: OnceLock<Arc<Mutex<Processor>>> = OnceLock::new();
+
+fn get_or_init_processor() -> anyhow::Result<Arc<Mutex<Processor>>> {
+    if let Some(p) = PROCESSOR.get() {
+        return Ok(p.clone());
+    }
+    let init_config = InitializationConfig {
+        num_capture_channels: APM_NUM_CHANNELS as i32,
+        num_render_channels: APM_NUM_CHANNELS as i32,
+        ..Default::default()
+    };
+    let mut processor = Processor::new(&init_config)
+        .map_err(|e| anyhow!("webrtc APM init failed: {e:?}"))?;
+
+    let config = Config {
+        echo_cancellation: Some(EchoCancellation {
+            suppression_level: EchoCancellationSuppressionLevel::High,
+            stream_delay_ms: Some(STREAM_DELAY_MS),
+            enable_delay_agnostic: true,
+            enable_extended_filter: true,
+        }),
+        noise_suppression: Some(NoiseSuppression {
+            suppression_level: NoiseSuppressionLevel::High,
+        }),
+        enable_high_pass_filter: true,
+        // AGC left off for now — it can fight the Opus encoder's own gain
+        // staging and the adaptive-quality controller. Add later if users
+        // report low mic levels.
+        ..Default::default()
+    };
+    processor.set_config(config);
+
+    let arc = Arc::new(Mutex::new(processor));
+    let _ = PROCESSOR.set(arc.clone());
+    info!(
+        stream_delay_ms = STREAM_DELAY_MS,
+        "webrtc APM initialized (AEC High + NS High + HPF, AGC off)"
+    );
+    Ok(arc)
+}
+
+// ---------------------------------------------------------------------------
+// Helpers: i16 ↔ f32 and APM frame processing
+// ---------------------------------------------------------------------------
+
+#[inline]
+fn i16_to_f32(s: i16) -> f32 {
+    s as f32 / 32768.0
+}
+
+#[inline]
+fn f32_to_i16(s: f32) -> i16 {
+    (s.clamp(-1.0, 1.0) * 32767.0) as i16
+}
+
+/// Feed a 20 ms (960-sample) playback frame to APM as the render reference.
+/// Splits into two 10 ms halves because APM is strict about frame size.
+/// Takes the Mutex-wrapped Processor and locks briefly around each call.
+fn push_render_frame_20ms(apm: &Mutex<Processor>, pcm: &[i16]) {
+    debug_assert_eq!(pcm.len(), FRAME_SAMPLES);
+    let mut buf = [0f32; APM_FRAME_SAMPLES];
+    for half in pcm.chunks_exact(APM_FRAME_SAMPLES) {
+        for (i, &s) in half.iter().enumerate() {
+            buf[i] = i16_to_f32(s);
+        }
+        match apm.lock() {
+            Ok(mut p) => {
+                if let Err(e) = p.process_render_frame(&mut buf) {
+                    warn!("webrtc APM process_render_frame failed: {e:?}");
+                }
+            }
+            Err(_) => {
+                warn!("webrtc APM mutex poisoned in render path");
+                return;
+            }
+        }
+    }
+}
+
+/// Run a 20 ms (960-sample) capture frame through APM's echo cancellation
+/// in place. Splits into two 10 ms halves, runs APM on each, stitches
+/// results back into the caller's buffer. Briefly holds the Mutex once
+/// per 10 ms half.
+fn process_capture_frame_20ms(apm: &Mutex<Processor>, pcm: &mut [i16]) {
+    debug_assert_eq!(pcm.len(), FRAME_SAMPLES);
+    let mut buf = [0f32; APM_FRAME_SAMPLES];
+    for half in pcm.chunks_exact_mut(APM_FRAME_SAMPLES) {
+        for (i, &s) in half.iter().enumerate() {
+            buf[i] = i16_to_f32(s);
+        }
+        match apm.lock() {
+            Ok(mut p) => {
+                if let Err(e) = p.process_capture_frame(&mut buf) {
+                    warn!("webrtc APM process_capture_frame failed: {e:?}");
+                }
+            }
+            Err(_) => {
+                warn!("webrtc APM mutex poisoned in capture path");
+                return;
+            }
+        }
+        for (i, d) in half.iter_mut().enumerate() {
+            *d = f32_to_i16(buf[i]);
+        }
+    }
+}
+
+// ---------------------------------------------------------------------------
+// LinuxAecCapture — CPAL mic + WebRTC AEC capture-side processing
+// ---------------------------------------------------------------------------
+
+/// Microphone capture with WebRTC AEC3 applied in place before the codec
+/// sees the samples. Mirrors the public API of `audio_io::AudioCapture` so
+/// downstream code doesn't change.
+pub struct LinuxAecCapture {
+    ring: Arc<AudioRing>,
+    running: Arc<AtomicBool>,
+}
+
+impl LinuxAecCapture {
+    pub fn start() -> Result<Self, anyhow::Error> {
+        // Eagerly init the APM so the playback side can find it already
+        // configured, and so init errors surface on the caller thread
+        // instead of silently failing inside the capture thread.
+        let apm = get_or_init_processor()?;
+
+        let ring = Arc::new(AudioRing::new());
+        let running = Arc::new(AtomicBool::new(true));
+
+        let (init_tx, init_rx) = std::sync::mpsc::sync_channel::<Result<(), String>>(1);
+
+        let ring_cb = ring.clone();
+        let running_clone = running.clone();
+        let apm_capture = apm.clone();
+
+        std::thread::Builder::new()
+            .name("wzp-audio-capture-linuxaec".into())
+            .spawn(move || {
+                let result = (|| -> Result<(), anyhow::Error> {
+                    let host = cpal::default_host();
+                    let device = host
+                        .default_input_device()
+                        .ok_or_else(|| anyhow!("no default input audio device found"))?;
+                    info!(device = %device.name().unwrap_or_default(), "LinuxAEC: using input device");
+
+                    let config = StreamConfig {
+                        channels: 1,
+                        sample_rate: SampleRate(48_000),
+                        buffer_size: cpal::BufferSize::Default,
+                    };
+
+                    let use_f32 = !supports_i16_input(&device)?;
+
+                    let err_cb = |e: cpal::StreamError| {
+                        warn!("LinuxAEC input stream error: {e}");
+                    };
+
+                    // Leftover buffer for when CPAL gives us partial frames.
+                    // We need exactly 960-sample chunks to feed APM.
+                    let leftover = std::sync::Mutex::new(Vec::<i16>::with_capacity(FRAME_SAMPLES * 4));
+
+                    let stream = if use_f32 {
+                        let ring = ring_cb.clone();
+                        let running = running_clone.clone();
+                        let apm = apm_capture.clone();
+                        device.build_input_stream(
+                            &config,
+                            move |data: &[f32], _: &cpal::InputCallbackInfo| {
+                                if !running.load(Ordering::Relaxed) {
+                                    return;
+                                }
+                                let mut lv = leftover.lock().unwrap();
+                                lv.reserve(data.len());
+                                for &s in data {
+                                    lv.push(f32_to_i16(s));
+                                }
+                                drain_frames_through_apm(&mut lv, &apm, &ring);
+                            },
+                            err_cb,
+                            None,
+                        )?
+                    } else {
+                        let ring = ring_cb.clone();
+                        let running = running_clone.clone();
+                        let apm = apm_capture.clone();
+                        device.build_input_stream(
+                            &config,
+                            move |data: &[i16], _: &cpal::InputCallbackInfo| {
+                                if !running.load(Ordering::Relaxed) {
+                                    return;
+                                }
+                                let mut lv = leftover.lock().unwrap();
+                                lv.extend_from_slice(data);
+                                drain_frames_through_apm(&mut lv, &apm, &ring);
+                            },
+                            err_cb,
+                            None,
+                        )?
+                    };
+
+                    stream.play().context("failed to start LinuxAEC input stream")?;
+                    let _ = init_tx.send(Ok(()));
+                    info!("LinuxAEC capture started (AEC3 active)");
+
+                    while running_clone.load(Ordering::Relaxed) {
+                        std::thread::park_timeout(std::time::Duration::from_millis(200));
+                    }
+                    drop(stream);
+                    Ok(())
+                })();
+
+                if let Err(e) = result {
+                    let _ = init_tx.send(Err(e.to_string()));
+                }
+            })?;
+
+        init_rx
+            .recv()
+            .map_err(|_| anyhow!("LinuxAEC capture thread exited before signaling"))?
+            .map_err(|e| anyhow!("{e}"))?;
+
+        Ok(Self { ring, running })
+    }
+
+    pub fn ring(&self) -> &Arc<AudioRing> {
+        &self.ring
+    }
+
+    pub fn stop(&self) {
+        self.running.store(false, Ordering::Relaxed);
+    }
+}
+
+impl Drop for LinuxAecCapture {
+    fn drop(&mut self) {
+        self.stop();
+    }
+}
+
+/// Pull whole 960-sample frames out of the leftover buffer, run them through
+/// APM's capture-side processing, and push to the ring. Leaves any partial
+/// sub-960 remainder in `leftover` for the next callback.
+fn drain_frames_through_apm(leftover: &mut Vec<i16>, apm: &Mutex<Processor>, ring: &AudioRing) {
+    let mut frame = [0i16; FRAME_SAMPLES];
+    while leftover.len() >= FRAME_SAMPLES {
+        frame.copy_from_slice(&leftover[..FRAME_SAMPLES]);
+        process_capture_frame_20ms(apm, &mut frame);
+        ring.write(&frame);
+        leftover.drain(..FRAME_SAMPLES);
+    }
+}
+
+// ---------------------------------------------------------------------------
+// LinuxAecPlayback — CPAL speaker output + WebRTC AEC render-side tee
+// ---------------------------------------------------------------------------
+
+/// Speaker playback with a render-side tee: each frame written to CPAL is
+/// ALSO fed to APM via `process_render_frame` as the echo-cancellation
+/// reference signal. This is the "tee the playback ring" approach (Zoom,
+/// Teams, Jitsi) — deterministic, does not depend on PulseAudio loopback or
+/// PipeWire monitor sources.
+pub struct LinuxAecPlayback {
+    ring: Arc<AudioRing>,
+    running: Arc<AtomicBool>,
+}
+
+impl LinuxAecPlayback {
+    pub fn start() -> Result<Self, anyhow::Error> {
+        let apm = get_or_init_processor()?;
+
+        let ring = Arc::new(AudioRing::new());
+        let running = Arc::new(AtomicBool::new(true));
+
+        let (init_tx, init_rx) = std::sync::mpsc::sync_channel::<Result<(), String>>(1);
+
+        let ring_cb = ring.clone();
+        let running_clone = running.clone();
+        let apm_render = apm.clone();
+
+        std::thread::Builder::new()
+            .name("wzp-audio-playback-linuxaec".into())
+            .spawn(move || {
+                let result = (|| -> Result<(), anyhow::Error> {
+                    let host = cpal::default_host();
+                    let device = host
+                        .default_output_device()
+                        .ok_or_else(|| anyhow!("no default output audio device found"))?;
+                    info!(device = %device.name().unwrap_or_default(), "LinuxAEC: using output device");
+
+                    let config = StreamConfig {
+                        channels: 1,
+                        sample_rate: SampleRate(48_000),
+                        buffer_size: cpal::BufferSize::Default,
+                    };
+
+                    let use_f32 = !supports_i16_output(&device)?;
+
+                    let err_cb = |e: cpal::StreamError| {
+                        warn!("LinuxAEC output stream error: {e}");
+                    };
+
+                    // Same 960-sample batching approach as the capture side:
+                    // CPAL may ask for N samples in a callback where N doesn't
+                    // divide 960. We accumulate partial frames in a Vec and
+                    // feed APM as soon as we have a whole 20 ms frame.
+                    let carry = std::sync::Mutex::new(Vec::<i16>::with_capacity(FRAME_SAMPLES * 4));
+
+                    let stream = if use_f32 {
+                        let ring = ring_cb.clone();
+                        let apm = apm_render.clone();
+                        device.build_output_stream(
+                            &config,
+                            move |data: &mut [f32], _: &cpal::OutputCallbackInfo| {
+                                fill_output_and_tee_f32(data, &ring, &apm, &carry);
+                            },
+                            err_cb,
+                            None,
+                        )?
+                    } else {
+                        let ring = ring_cb.clone();
+                        let apm = apm_render.clone();
+                        device.build_output_stream(
+                            &config,
+                            move |data: &mut [i16], _: &cpal::OutputCallbackInfo| {
+                                fill_output_and_tee_i16(data, &ring, &apm, &carry);
+                            },
+                            err_cb,
+                            None,
+                        )?
+                    };
+
+                    stream.play().context("failed to start LinuxAEC output stream")?;
+                    let _ = init_tx.send(Ok(()));
+                    info!("LinuxAEC playback started (render tee active)");
+
+                    while running_clone.load(Ordering::Relaxed) {
+                        std::thread::park_timeout(std::time::Duration::from_millis(200));
+                    }
+                    drop(stream);
+                    Ok(())
+                })();
+
+                if let Err(e) = result {
+                    let _ = init_tx.send(Err(e.to_string()));
+                }
+            })?;
+
+        init_rx
+            .recv()
+            .map_err(|_| anyhow!("LinuxAEC playback thread exited before signaling"))?
+            .map_err(|e| anyhow!("{e}"))?;
+
+        Ok(Self { ring, running })
+    }
+
+    pub fn ring(&self) -> &Arc<AudioRing> {
+        &self.ring
+    }
+
+    pub fn stop(&self) {
+        self.running.store(false, Ordering::Relaxed);
+    }
+}
+
+impl Drop for LinuxAecPlayback {
+    fn drop(&mut self) {
+        self.stop();
+    }
+}
+
+fn fill_output_and_tee_i16(
+    data: &mut [i16],
+    ring: &AudioRing,
+    apm: &Mutex<Processor>,
+    carry: &std::sync::Mutex<Vec<i16>>,
+) {
+    let read = ring.read(data);
+    for s in &mut data[read..] {
+        *s = 0;
+    }
+    tee_render_samples(data, apm, carry);
+}
+
+fn fill_output_and_tee_f32(
+    data: &mut [f32],
+    ring: &AudioRing,
+    apm: &Mutex<Processor>,
+    carry: &std::sync::Mutex<Vec<i16>>,
+) {
+    let mut tmp = vec![0i16; data.len()];
+    let read = ring.read(&mut tmp);
+    for s in &mut tmp[read..] {
+        *s = 0;
+    }
+    for (d, &s) in data.iter_mut().zip(tmp.iter()) {
+        *d = i16_to_f32(s);
+    }
+    tee_render_samples(&tmp, apm, carry);
+}
+
+/// Push CPAL-bound samples into APM's render-side input for echo cancellation.
+/// Uses a carry buffer to batch into exact 960-sample (20 ms) frames.
+fn tee_render_samples(samples: &[i16], apm: &Mutex<Processor>, carry: &std::sync::Mutex<Vec<i16>>) {
+    let mut lv = carry.lock().unwrap();
+    lv.extend_from_slice(samples);
+    while lv.len() >= FRAME_SAMPLES {
+        let mut frame = [0i16; FRAME_SAMPLES];
+        frame.copy_from_slice(&lv[..FRAME_SAMPLES]);
+        push_render_frame_20ms(apm, &frame);
+        lv.drain(..FRAME_SAMPLES);
+    }
+}
+
+// ---------------------------------------------------------------------------
+// CPAL format helpers (duplicated from audio_io.rs to keep the modules
+// independent — each backend file is a self-contained unit)
+// ---------------------------------------------------------------------------
+
+fn supports_i16_input(device: &cpal::Device) -> Result<bool, anyhow::Error> {
+    let supported = device
+        .supported_input_configs()
+        .context("failed to query input configs")?;
+    for cfg in supported {
+        if cfg.sample_format() == SampleFormat::I16
+            && cfg.min_sample_rate() <= SampleRate(48_000)
+            && cfg.max_sample_rate() >= SampleRate(48_000)
+            && cfg.channels() >= 1
+        {
+            return Ok(true);
+        }
+    }
+    Ok(false)
+}
+
+fn supports_i16_output(device: &cpal::Device) -> Result<bool, anyhow::Error> {
+    let supported = device
+        .supported_output_configs()
+        .context("failed to query output configs")?;
+    for cfg in supported {
+        if cfg.sample_format() == SampleFormat::I16
+            && cfg.min_sample_rate() <= SampleRate(48_000)
+            && cfg.max_sample_rate() >= SampleRate(48_000)
+            && cfg.channels() >= 1
+        {
+            return Ok(true);
+        }
+    }
+    Ok(false)
+}

crates/wzp-client/src/audio_ring.rs

@@ -0,0 +1,122 @@
+//! Lock-free SPSC ring buffer — "Reader-Detects-Lap" architecture.
+//!
+//! SPSC invariant: the producer ONLY writes `write_pos`, the consumer
+//! ONLY writes `read_pos`.  Neither thread touches the other's cursor.
+//!
+//! On overflow (writer laps the reader), the writer simply overwrites
+//! old buffer data.  The reader detects the lap via `available() >
+//! RING_CAPACITY` and snaps its own `read_pos` forward.
+//!
+//! Capacity is a power of 2 for bitmask indexing (no modulo).
+
+use std::sync::atomic::{AtomicU64, AtomicUsize, Ordering};
+
+/// Ring buffer capacity — power of 2 for bitmask indexing.
+/// 16384 samples = 341.3ms at 48kHz mono.
+const RING_CAPACITY: usize = 16384; // 2^14
+const RING_MASK: usize = RING_CAPACITY - 1;
+
+/// Lock-free single-producer single-consumer ring buffer for i16 PCM samples.
+pub struct AudioRing {
+    buf: Box<[i16]>,
+    /// Monotonically increasing write cursor. ONLY written by producer.
+    write_pos: AtomicUsize,
+    /// Monotonically increasing read cursor. ONLY written by consumer.
+    read_pos: AtomicUsize,
+    /// Incremented by reader when it detects it was lapped (overflow).
+    overflow_count: AtomicU64,
+    /// Incremented by reader when ring is empty (underrun).
+    underrun_count: AtomicU64,
+}
+
+// SAFETY: AudioRing is SPSC — one thread writes (producer), one reads (consumer).
+// The producer only writes write_pos. The consumer only writes read_pos.
+// Neither thread writes the other's cursor. Buffer indices are derived from
+// the owning thread's cursor, ensuring no concurrent access to the same index.
+unsafe impl Send for AudioRing {}
+unsafe impl Sync for AudioRing {}
+
+impl AudioRing {
+    pub fn new() -> Self {
+        debug_assert!(RING_CAPACITY.is_power_of_two());
+        Self {
+            buf: vec![0i16; RING_CAPACITY].into_boxed_slice(),
+            write_pos: AtomicUsize::new(0),
+            read_pos: AtomicUsize::new(0),
+            overflow_count: AtomicU64::new(0),
+            underrun_count: AtomicU64::new(0),
+        }
+    }
+
+    /// Number of samples available to read (clamped to capacity).
+    pub fn available(&self) -> usize {
+        let w = self.write_pos.load(Ordering::Acquire);
+        let r = self.read_pos.load(Ordering::Relaxed);
+        w.wrapping_sub(r).min(RING_CAPACITY)
+    }
+
+    /// Write samples into the ring. Returns number of samples written.
+    ///
+    /// If the ring is full, old data is silently overwritten.  The reader
+    /// will detect the lap and self-correct.  The writer NEVER touches
+    /// `read_pos`.
+    pub fn write(&self, samples: &[i16]) -> usize {
+        let count = samples.len().min(RING_CAPACITY);
+        let w = self.write_pos.load(Ordering::Relaxed);
+
+        for i in 0..count {
+            unsafe {
+                let ptr = self.buf.as_ptr() as *mut i16;
+                *ptr.add((w + i) & RING_MASK) = samples[i];
+            }
+        }
+
+        self.write_pos
+            .store(w.wrapping_add(count), Ordering::Release);
+        count
+    }
+
+    /// Read samples from the ring into `out`. Returns number of samples read.
+    ///
+    /// If the writer has lapped the reader (overflow), `read_pos` is snapped
+    /// forward to the oldest valid data.
+    pub fn read(&self, out: &mut [i16]) -> usize {
+        let w = self.write_pos.load(Ordering::Acquire);
+        let mut r = self.read_pos.load(Ordering::Relaxed);
+
+        let mut avail = w.wrapping_sub(r);
+
+        // Lap detection: writer has overwritten our unread data.
+        if avail > RING_CAPACITY {
+            r = w.wrapping_sub(RING_CAPACITY);
+            avail = RING_CAPACITY;
+            self.overflow_count.fetch_add(1, Ordering::Relaxed);
+        }
+
+        let count = out.len().min(avail);
+        if count == 0 {
+            if w == r {
+                self.underrun_count.fetch_add(1, Ordering::Relaxed);
+            }
+            return 0;
+        }
+
+        for i in 0..count {
+            out[i] = unsafe { *self.buf.as_ptr().add((r + i) & RING_MASK) };
+        }
+
+        self.read_pos
+            .store(r.wrapping_add(count), Ordering::Release);
+        count
+    }
+
+    /// Number of overflow events (reader was lapped by writer).
+    pub fn overflow_count(&self) -> u64 {
+        self.overflow_count.load(Ordering::Relaxed)
+    }
+
+    /// Number of underrun events (reader found empty buffer).
+    pub fn underrun_count(&self) -> u64 {
+        self.underrun_count.load(Ordering::Relaxed)
+    }
+}

crates/wzp-client/src/audio_vpio.rs

@@ -0,0 +1,179 @@
+//! macOS Voice Processing I/O — uses Apple's VoiceProcessingIO audio unit
+//! for hardware-accelerated echo cancellation, AGC, and noise suppression.
+//!
+//! VoiceProcessingIO is a combined input+output unit that knows what's going
+//! to the speaker, so it can cancel the echo from the mic signal internally.
+//! This is the same engine FaceTime and other Apple apps use.
+
+use std::sync::atomic::{AtomicBool, Ordering};
+use std::sync::Arc;
+
+use anyhow::Context;
+use coreaudio::audio_unit::audio_format::LinearPcmFlags;
+use coreaudio::audio_unit::render_callback::{self, data};
+use coreaudio::audio_unit::{AudioUnit, Element, IOType, SampleFormat, Scope, StreamFormat};
+use coreaudio::sys;
+use tracing::info;
+
+use crate::audio_ring::AudioRing;
+
+/// Number of samples per 20 ms frame at 48 kHz mono.
+pub const FRAME_SAMPLES: usize = 960;
+
+/// Combined capture + playback via macOS VoiceProcessingIO.
+///
+/// The OS handles AEC internally — no manual far-end feeding needed.
+pub struct VpioAudio {
+    capture_ring: Arc<AudioRing>,
+    playout_ring: Arc<AudioRing>,
+    _audio_unit: AudioUnit,
+    running: Arc<AtomicBool>,
+}
+
+impl VpioAudio {
+    /// Start VoiceProcessingIO with AEC enabled.
+    pub fn start() -> Result<Self, anyhow::Error> {
+        let capture_ring = Arc::new(AudioRing::new());
+        let playout_ring = Arc::new(AudioRing::new());
+        let running = Arc::new(AtomicBool::new(true));
+
+        let mut au = AudioUnit::new(IOType::VoiceProcessingIO)
+            .context("failed to create VoiceProcessingIO audio unit")?;
+
+        // Must uninitialize before configuring properties.
+        au.uninitialize()
+            .context("failed to uninitialize VPIO for configuration")?;
+
+        // Enable input (mic) on Element::Input (bus 1).
+        let enable: u32 = 1;
+        au.set_property(
+            sys::kAudioOutputUnitProperty_EnableIO,
+            Scope::Input,
+            Element::Input,
+            Some(&enable),
+        )
+        .context("failed to enable VPIO input")?;
+
+        // Output (speaker) is enabled by default on VPIO, but be explicit.
+        au.set_property(
+            sys::kAudioOutputUnitProperty_EnableIO,
+            Scope::Output,
+            Element::Output,
+            Some(&enable),
+        )
+        .context("failed to enable VPIO output")?;
+
+        // Configure stream format: 48kHz mono f32 non-interleaved
+        let stream_format = StreamFormat {
+            sample_rate: 48_000.0,
+            sample_format: SampleFormat::F32,
+            flags: LinearPcmFlags::IS_FLOAT
+                | LinearPcmFlags::IS_PACKED
+                | LinearPcmFlags::IS_NON_INTERLEAVED,
+            channels: 1,
+        };
+
+        let asbd = stream_format.to_asbd();
+
+        // Input: set format on Output scope of Input element
+        // (= the format the AU delivers to us from the mic)
+        au.set_property(
+            sys::kAudioUnitProperty_StreamFormat,
+            Scope::Output,
+            Element::Input,
+            Some(&asbd),
+        )
+        .context("failed to set input stream format")?;
+
+        // Output: set format on Input scope of Output element
+        // (= the format we feed to the AU for the speaker)
+        au.set_property(
+            sys::kAudioUnitProperty_StreamFormat,
+            Scope::Input,
+            Element::Output,
+            Some(&asbd),
+        )
+        .context("failed to set output stream format")?;
+
+        // Set up input callback (mic capture with AEC applied)
+        let cap_ring = capture_ring.clone();
+        let cap_running = running.clone();
+        let logged = Arc::new(AtomicBool::new(false));
+        au.set_input_callback(
+            move |args: render_callback::Args<data::NonInterleaved<f32>>| {
+                if !cap_running.load(Ordering::Relaxed) {
+                    return Ok(());
+                }
+                let mut buffers = args.data.channels();
+                if let Some(ch) = buffers.next() {
+                    if !logged.swap(true, Ordering::Relaxed) {
+                        eprintln!("[vpio] capture callback: {} f32 samples", ch.len());
+                    }
+                    let mut tmp = [0i16; FRAME_SAMPLES];
+                    for chunk in ch.chunks(FRAME_SAMPLES) {
+                        let n = chunk.len();
+                        for i in 0..n {
+                            tmp[i] = (chunk[i].clamp(-1.0, 1.0) * i16::MAX as f32) as i16;
+                        }
+                        cap_ring.write(&tmp[..n]);
+                    }
+                }
+                Ok(())
+            },
+        )
+        .context("failed to set input callback")?;
+
+        // Set up output callback (speaker playback — AEC uses this as reference)
+        let play_ring = playout_ring.clone();
+        au.set_render_callback(
+            move |mut args: render_callback::Args<data::NonInterleaved<f32>>| {
+                let mut buffers = args.data.channels_mut();
+                if let Some(ch) = buffers.next() {
+                    let mut tmp = [0i16; FRAME_SAMPLES];
+                    for chunk in ch.chunks_mut(FRAME_SAMPLES) {
+                        let n = chunk.len();
+                        let read = play_ring.read(&mut tmp[..n]);
+                        for i in 0..read {
+                            chunk[i] = tmp[i] as f32 / i16::MAX as f32;
+                        }
+                        for i in read..n {
+                            chunk[i] = 0.0;
+                        }
+                    }
+                }
+                Ok(())
+            },
+        )
+        .context("failed to set render callback")?;
+
+        au.initialize().context("failed to initialize VoiceProcessingIO")?;
+        au.start().context("failed to start VoiceProcessingIO")?;
+
+        info!("VoiceProcessingIO started (OS-level AEC enabled)");
+
+        Ok(Self {
+            capture_ring,
+            playout_ring,
+            _audio_unit: au,
+            running,
+        })
+    }
+
+    pub fn capture_ring(&self) -> &Arc<AudioRing> {
+        &self.capture_ring
+    }
+
+    pub fn playout_ring(&self) -> &Arc<AudioRing> {
+        &self.playout_ring
+    }
+
+    pub fn stop(&self) {
+        self.running.store(false, Ordering::Relaxed);
+    }
+}
+
+impl Drop for VpioAudio {
+    fn drop(&mut self) {
+        self.stop();
+    }
+}

crates/wzp-client/src/audio_wasapi.rs

@@ -0,0 +1,332 @@
+//! Direct WASAPI microphone capture with Windows's OS-level AEC enabled.
+//!
+//! Bypasses CPAL and opens the default capture endpoint directly via
+//! `IMMDeviceEnumerator` + `IAudioClient2::SetClientProperties`, setting
+//! `AudioClientProperties.eCategory = AudioCategory_Communications`. That's
+//! the switch that tells Windows "this is a VoIP call" — the OS then
+//! enables its communications audio processing chain (AEC, noise
+//! suppression, automatic gain control) for the stream. AEC operates at
+//! the OS level using the currently-playing audio as the reference
+//! signal, so it cancels echo from our CPAL playback (and any other app's
+//! audio) without us having to plumb a reference signal ourselves.
+//!
+//! Platform: Windows only, compiled only when the `windows-aec` feature
+//! is enabled. Mirrors the public API of `audio_io::AudioCapture` so
+//! `wzp-client`'s lib.rs can transparently re-export either one as
+//! `AudioCapture`.
+
+use std::sync::atomic::{AtomicBool, Ordering};
+use std::sync::Arc;
+
+use anyhow::{anyhow, Context};
+use tracing::{info, warn};
+use windows::core::{Interface, GUID};
+use windows::Win32::Foundation::{CloseHandle, BOOL, WAIT_OBJECT_0};
+use windows::Win32::Media::Audio::{
+    eCapture, eCommunications, AudioCategory_Communications, AudioClientProperties,
+    IAudioCaptureClient, IAudioClient, IAudioClient2, IMMDeviceEnumerator, MMDeviceEnumerator,
+    AUDCLNT_SHAREMODE_SHARED, AUDCLNT_STREAMFLAGS_AUTOCONVERTPCM,
+    AUDCLNT_STREAMFLAGS_EVENTCALLBACK, AUDCLNT_STREAMFLAGS_SRC_DEFAULT_QUALITY, WAVEFORMATEX,
+    WAVE_FORMAT_PCM,
+};
+use windows::Win32::System::Com::{
+    CoCreateInstance, CoInitializeEx, CoUninitialize, CLSCTX_ALL, COINIT_MULTITHREADED,
+};
+use windows::Win32::System::Threading::{CreateEventW, WaitForSingleObject, INFINITE};
+
+use crate::audio_ring::AudioRing;
+
+/// 20 ms at 48 kHz, mono. Matches the rest of the audio pipeline.
+pub const FRAME_SAMPLES: usize = 960;
+
+/// Microphone capture via WASAPI with Windows's communications AEC enabled.
+///
+/// The WASAPI capture stream runs on a dedicated OS thread. This handle is
+/// `Send + Sync`. Dropping it stops the stream and joins the thread.
+pub struct WasapiAudioCapture {
+    ring: Arc<AudioRing>,
+    running: Arc<AtomicBool>,
+    thread: Option<std::thread::JoinHandle<()>>,
+}
+
+impl WasapiAudioCapture {
+    /// Open the default communications microphone, enable OS AEC, and start
+    /// streaming PCM into a lock-free ring buffer.
+    ///
+    /// Returns only after the capture thread has successfully initialized
+    /// the stream, or propagates the error back to the caller.
+    pub fn start() -> Result<Self, anyhow::Error> {
+        let ring = Arc::new(AudioRing::new());
+        let running = Arc::new(AtomicBool::new(true));
+
+        let (init_tx, init_rx) = std::sync::mpsc::sync_channel::<Result<(), String>>(1);
+        let ring_cb = ring.clone();
+        let running_cb = running.clone();
+
+        let thread = std::thread::Builder::new()
+            .name("wzp-audio-capture-wasapi".into())
+            .spawn(move || {
+                let result = unsafe { capture_thread_main(ring_cb, running_cb.clone(), &init_tx) };
+                if let Err(e) = result {
+                    warn!("wasapi capture thread exited with error: {e}");
+                    // If we failed before signaling init, signal now so the
+                    // caller unblocks. Double-send is harmless (channel is
+                    // bounded to 1 and we only hit the second send path on
+                    // late errors).
+                    let _ = init_tx.send(Err(e.to_string()));
+                }
+            })
+            .context("failed to spawn WASAPI capture thread")?;
+
+        init_rx
+            .recv()
+            .map_err(|_| anyhow!("WASAPI capture thread exited before signaling init"))?
+            .map_err(|e| anyhow!("{e}"))?;
+
+        Ok(Self {
+            ring,
+            running,
+            thread: Some(thread),
+        })
+    }
+
+    /// Get a reference to the capture ring buffer for direct polling.
+    pub fn ring(&self) -> &Arc<AudioRing> {
+        &self.ring
+    }
+
+    /// Stop capturing.
+    pub fn stop(&self) {
+        self.running.store(false, Ordering::Relaxed);
+    }
+}
+
+impl Drop for WasapiAudioCapture {
+    fn drop(&mut self) {
+        self.stop();
+        if let Some(handle) = self.thread.take() {
+            // Join best-effort. The thread loop polls `running` every 200ms
+            // via a short WaitForSingleObject timeout, so it should exit
+            // within ~200ms of `stop()`.
+            let _ = handle.join();
+        }
+    }
+}
+
+// ---------------------------------------------------------------------------
+// WASAPI thread entry point — everything below this line runs on the
+// dedicated wzp-audio-capture-wasapi thread.
+// ---------------------------------------------------------------------------
+
+unsafe fn capture_thread_main(
+    ring: Arc<AudioRing>,
+    running: Arc<AtomicBool>,
+    init_tx: &std::sync::mpsc::SyncSender<Result<(), String>>,
+) -> Result<(), anyhow::Error> {
+    // COM init for the capture thread. MULTITHREADED because we're not
+    // running a message pump. Must be balanced by CoUninitialize on exit.
+    CoInitializeEx(None, COINIT_MULTITHREADED)
+        .ok()
+        .context("CoInitializeEx failed")?;
+
+    // Use a guard struct so CoUninitialize runs even on early returns.
+    struct ComGuard;
+    impl Drop for ComGuard {
+        fn drop(&mut self) {
+            unsafe { CoUninitialize() };
+        }
+    }
+    let _com_guard = ComGuard;
+
+    let enumerator: IMMDeviceEnumerator =
+        CoCreateInstance(&MMDeviceEnumerator, None, CLSCTX_ALL)
+            .context("CoCreateInstance(MMDeviceEnumerator) failed")?;
+
+    // eCommunications role (not eConsole) — this picks the device the user
+    // has designated for communications in Sound Settings. It's the one
+    // Windows's AEC is actually tuned for and the one Teams/Zoom use.
+    let device = enumerator
+        .GetDefaultAudioEndpoint(eCapture, eCommunications)
+        .context("GetDefaultAudioEndpoint(eCapture, eCommunications) failed")?;
+
+    if let Ok(name) = device_name(&device) {
+        info!(device = %name, "opening WASAPI communications capture endpoint");
+    }
+
+    let audio_client: IAudioClient = device
+        .Activate(CLSCTX_ALL, None)
+        .context("IMMDevice::Activate(IAudioClient) failed")?;
+
+    // IAudioClient2 exposes SetClientProperties, which is the ONLY way to
+    // set AudioCategory_Communications pre-Initialize. Calling it on the
+    // base IAudioClient would not compile, and setting it after Initialize
+    // is a no-op.
+    let audio_client2: IAudioClient2 = audio_client
+        .cast()
+        .context("QueryInterface IAudioClient2 failed")?;
+
+    let mut props = AudioClientProperties {
+        cbSize: std::mem::size_of::<AudioClientProperties>() as u32,
+        bIsOffload: BOOL(0),
+        eCategory: AudioCategory_Communications,
+        // 0 = AUDCLNT_STREAMOPTIONS_NONE. The `windows` crate doesn't
+        // export the enum constant in all versions, so use 0 directly.
+        Options: Default::default(),
+    };
+    audio_client2
+        .SetClientProperties(&mut props as *mut _)
+        .context("SetClientProperties(AudioCategory_Communications) failed")?;
+
+    // Request 48 kHz mono i16 directly. AUDCLNT_STREAMFLAGS_AUTOCONVERTPCM
+    // tells Windows to do any needed format conversion inside the audio
+    // engine rather than rejecting our format. SRC_DEFAULT_QUALITY picks
+    // the standard Windows resampler quality (fine for voice).
+    let wave_format = WAVEFORMATEX {
+        wFormatTag: WAVE_FORMAT_PCM as u16,
+        nChannels: 1,
+        nSamplesPerSec: 48_000,
+        nAvgBytesPerSec: 48_000 * 2, // 1 ch * 2 bytes/sample * 48000 Hz
+        nBlockAlign: 2,              // 1 ch * 2 bytes/sample
+        wBitsPerSample: 16,
+        cbSize: 0,
+    };
+
+    // 1,000,000 hns = 100 ms buffer (hns = 100-nanosecond units). Windows
+    // treats this as the minimum; the engine may give us a larger one.
+    const BUFFER_DURATION_HNS: i64 = 1_000_000;
+
+    audio_client
+        .Initialize(
+            AUDCLNT_SHAREMODE_SHARED,
+            AUDCLNT_STREAMFLAGS_EVENTCALLBACK
+                | AUDCLNT_STREAMFLAGS_AUTOCONVERTPCM
+                | AUDCLNT_STREAMFLAGS_SRC_DEFAULT_QUALITY,
+            BUFFER_DURATION_HNS,
+            0,
+            &wave_format,
+            Some(&GUID::zeroed()),
+        )
+        .context("IAudioClient::Initialize failed — Windows rejected communications-mode 48k mono i16")?;
+
+    // Event-driven capture: Windows signals this handle each time a new
+    // audio packet is available. We wait on it from the loop below.
+    let event = CreateEventW(None, false, false, None)
+        .context("CreateEventW failed")?;
+    audio_client
+        .SetEventHandle(event)
+        .context("SetEventHandle failed")?;
+
+    let capture_client: IAudioCaptureClient = audio_client
+        .GetService()
+        .context("IAudioClient::GetService(IAudioCaptureClient) failed")?;
+
+    audio_client.Start().context("IAudioClient::Start failed")?;
+
+    // Signal to the parent thread that init succeeded before entering the
+    // hot loop. From this point on, errors get logged but don't propagate
+    // back to the caller (they'd just cause the ring buffer to stop
+    // filling, which the main thread detects as underruns).
+    let _ = init_tx.send(Ok(()));
+    info!("WASAPI communications-mode capture started with OS AEC enabled");
+
+    let mut logged_first_packet = false;
+
+    // Main capture loop. Exit when `running` goes false (from Drop or an
+    // explicit stop() call).
+    while running.load(Ordering::Relaxed) {
+        // 200 ms timeout so we check `running` regularly even if the audio
+        // engine stops delivering packets (e.g. device unplugged).
+        let wait = WaitForSingleObject(event, 200);
+        if wait.0 != WAIT_OBJECT_0.0 {
+            // Timeout or failure — just loop and re-check running.
+            continue;
+        }
+
+        // Drain all available packets. Windows may have queued more than
+        // one since we were last scheduled.
+        loop {
+            let packet_length = match capture_client.GetNextPacketSize() {
+                Ok(n) => n,
+                Err(e) => {
+                    warn!("GetNextPacketSize failed: {e}");
+                    break;
+                }
+            };
+            if packet_length == 0 {
+                break;
+            }
+
+            let mut buffer_ptr: *mut u8 = std::ptr::null_mut();
+            let mut num_frames: u32 = 0;
+            let mut flags: u32 = 0;
+            let mut device_position: u64 = 0;
+            let mut qpc_position: u64 = 0;
+
+            if let Err(e) = capture_client.GetBuffer(
+                &mut buffer_ptr,
+                &mut num_frames,
+                &mut flags,
+                Some(&mut device_position),
+                Some(&mut qpc_position),
+            ) {
+                warn!("GetBuffer failed: {e}");
+                break;
+            }
+
+            if num_frames > 0 && !buffer_ptr.is_null() {
+                if !logged_first_packet {
+                    info!(
+                        frames = num_frames,
+                        flags, "WASAPI capture: first packet received"
+                    );
+                    logged_first_packet = true;
+                }
+
+                // Because we asked for 48 kHz mono i16, each frame is
+                // exactly one i16. Windows's AUTOCONVERTPCM handles the
+                // conversion from whatever the engine mix format is.
+                let samples = std::slice::from_raw_parts(
+                    buffer_ptr as *const i16,
+                    num_frames as usize,
+                );
+                ring.write(samples);
+            }
+
+            if let Err(e) = capture_client.ReleaseBuffer(num_frames) {
+                warn!("ReleaseBuffer failed: {e}");
+                break;
+            }
+        }
+    }
+
+    info!("WASAPI capture thread stopping");
+    let _ = audio_client.Stop();
+    let _ = CloseHandle(event);
+    // _com_guard drops here, calling CoUninitialize.
+
+    // Silence INFINITE unused-import warning — it's referenced by the
+    // `windows` crate's WaitForSingleObject alternative but we use the
+    // 200 ms timeout variant instead. Explicit suppression for clarity.
+    let _ = INFINITE;
+
+    Ok(())
+}
+
+// ---------------------------------------------------------------------------
+// Helpers
+// ---------------------------------------------------------------------------
+
+/// Best-effort device ID string for logging. Grabbing the friendly name via
+/// PKEY_Device_FriendlyName requires IPropertyStore + PROPVARIANT plumbing
+/// that's far more ceremony than a log line justifies; the ID is already
+/// sufficient to confirm we opened the right endpoint.
+///
+/// Rust 2024 edition's `unsafe_op_in_unsafe_fn` lint requires explicit
+/// `unsafe { ... }` blocks inside `unsafe fn` bodies for each unsafe call,
+/// even though the whole function is already marked unsafe.
+unsafe fn device_name(
+    device: &windows::Win32::Media::Audio::IMMDevice,
+) -> Result<String, anyhow::Error> {
+    let id = unsafe { device.GetId() }.context("IMMDevice::GetId failed")?;
+    Ok(unsafe { id.to_string() }.unwrap_or_else(|_| "<non-utf16>".to_string()))
+}

crates/wzp-client/src/birthday.rs

@@ -0,0 +1,350 @@
+//! Birthday attack for hard NAT traversal.
+//!
+//! When both peers are behind symmetric NATs with random port
+//! allocation, standard hole-punching fails because neither side
+//! can predict the other's external port. This module implements
+//! the birthday-paradox approach:
+//!
+//! 1. **Acceptor** opens N sockets, STUN-probes each to learn
+//!    their external ports, reports them to the Dialer.
+//! 2. **Dialer** sprays QUIC connect attempts to the Acceptor's
+//!    reported ports + random ports on the Acceptor's IP.
+//! 3. Birthday paradox: with N=64 ports and M=256 probes across
+//!    65536 ports, collision probability is high.
+//!
+//! In practice, the Acceptor's STUN-probed ports are known
+//! exactly (not random), so the Dialer targets them first —
+//! making this more like "spray-and-pray with a hit list" than
+//! a pure birthday attack.
+
+use std::net::{Ipv4Addr, SocketAddr};
+use std::time::{Duration, Instant};
+
+use crate::stun;
+
+/// Configuration for the birthday attack.
+#[derive(Debug, Clone)]
+pub struct BirthdayConfig {
+    /// Number of sockets the Acceptor opens (default: 32).
+    /// Each socket gets STUN-probed to learn its external port.
+    /// More = higher chance of collision, but more resource usage.
+    pub acceptor_ports: u16,
+    /// Number of QUIC connect attempts the Dialer makes (default: 128).
+    /// Spread across the Acceptor's known ports + random ports.
+    pub dialer_probes: u16,
+    /// Rate limit: ms between consecutive probes (default: 20ms = 50/s).
+    pub probe_interval_ms: u16,
+    /// Overall timeout for the birthday attack phase.
+    pub timeout: Duration,
+    /// STUN config for probing external ports.
+    pub stun_config: stun::StunConfig,
+}
+
+impl Default for BirthdayConfig {
+    fn default() -> Self {
+        Self {
+            acceptor_ports: 32,
+            dialer_probes: 128,
+            probe_interval_ms: 20,
+            timeout: Duration::from_secs(8),
+            stun_config: stun::StunConfig {
+                servers: vec!["stun.l.google.com:19302".into()],
+                timeout: Duration::from_secs(2),
+            },
+        }
+    }
+}
+
+/// Result of the Acceptor's port-opening phase.
+#[derive(Debug, Clone, serde::Serialize)]
+pub struct AcceptorPorts {
+    /// External IP (from STUN).
+    pub external_ip: Option<Ipv4Addr>,
+    /// List of (local_port, external_port) for each opened socket.
+    pub ports: Vec<PortMapping>,
+    /// How many sockets we attempted to open.
+    pub attempted: u16,
+    /// How many STUN probes succeeded.
+    pub succeeded: u16,
+}
+
+/// A single socket's local↔external port mapping.
+#[derive(Debug, Clone, serde::Serialize)]
+pub struct PortMapping {
+    pub local_port: u16,
+    pub external_port: u16,
+}
+
+/// Open N sockets and STUN-probe each to discover external ports.
+///
+/// Returns the set of known external ports that the Dialer should
+/// target. Each socket stays open (bound) so the NAT mapping
+/// remains active until the returned `PortGuard` is dropped.
+///
+/// The sockets are returned so the caller can keep them alive
+/// during the attack. Dropping them closes the NAT pinholes.
+pub async fn open_acceptor_ports(
+    config: &BirthdayConfig,
+) -> (AcceptorPorts, Vec<tokio::net::UdpSocket>) {
+    let mut sockets = Vec::new();
+    let mut mappings = Vec::new();
+    let mut external_ip: Option<Ipv4Addr> = None;
+    let mut succeeded: u16 = 0;
+
+    let stun_server = match config.stun_config.servers.first() {
+        Some(s) => match stun::resolve_stun_server(s).await {
+            Ok(a) => Some(a),
+            Err(_) => None,
+        },
+        None => None,
+    };
+
+    for _ in 0..config.acceptor_ports {
+        // Bind to random port
+        let sock = match tokio::net::UdpSocket::bind("0.0.0.0:0").await {
+            Ok(s) => s,
+            Err(_) => continue,
+        };
+        let local_port = match sock.local_addr() {
+            Ok(a) => a.port(),
+            Err(_) => continue,
+        };
+
+        // STUN probe to learn external port
+        if let Some(stun_addr) = stun_server {
+            match stun::stun_reflect(&sock, stun_addr, config.stun_config.timeout).await {
+                Ok(ext_addr) => {
+                    if external_ip.is_none() {
+                        if let std::net::IpAddr::V4(ip) = ext_addr.ip() {
+                            external_ip = Some(ip);
+                        }
+                    }
+                    mappings.push(PortMapping {
+                        local_port,
+                        external_port: ext_addr.port(),
+                    });
+                    succeeded += 1;
+                }
+                Err(e) => {
+                    tracing::debug!(local_port, error = %e, "birthday: STUN probe failed for socket");
+                }
+            }
+        }
+
+        sockets.push(sock);
+    }
+
+    tracing::info!(
+        attempted = config.acceptor_ports,
+        succeeded,
+        external_ip = ?external_ip,
+        "birthday: acceptor ports opened"
+    );
+
+    let result = AcceptorPorts {
+        external_ip,
+        ports: mappings,
+        attempted: config.acceptor_ports,
+        succeeded,
+    };
+
+    (result, sockets)
+}
+
+/// Generate the list of target addresses for the Dialer to spray.
+///
+/// Priority order:
+/// 1. Acceptor's known external ports (from STUN probes) — highest hit rate
+/// 2. Random ports on the Acceptor's IP — birthday paradox fill
+pub fn generate_dialer_targets(
+    acceptor_ip: Ipv4Addr,
+    known_ports: &[u16],
+    total_probes: u16,
+) -> Vec<SocketAddr> {
+    let mut targets = Vec::with_capacity(total_probes as usize);
+
+    // First: all known ports (guaranteed targets)
+    for &port in known_ports {
+        targets.push(SocketAddr::new(
+            std::net::IpAddr::V4(acceptor_ip),
+            port,
+        ));
+    }
+
+    // Fill remaining with random ports (birthday attack)
+    let remaining = total_probes.saturating_sub(known_ports.len() as u16);
+    if remaining > 0 {
+        use rand::Rng;
+        let mut rng = rand::thread_rng();
+        for _ in 0..remaining {
+            let port = rng.gen_range(1024..=65535u16);
+            let addr = SocketAddr::new(
+                std::net::IpAddr::V4(acceptor_ip),
+                port,
+            );
+            if !targets.contains(&addr) {
+                targets.push(addr);
+            }
+        }
+    }
+
+    targets
+}
+
+/// Run the Dialer side of the birthday attack.
+///
+/// Sprays QUIC connection attempts at the target addresses.
+/// Returns the first successful connection, or None on timeout.
+pub async fn spray_dialer(
+    endpoint: &wzp_transport::Endpoint,
+    targets: &[SocketAddr],
+    call_sni: &str,
+    probe_interval: Duration,
+    timeout: Duration,
+) -> Option<wzp_transport::QuinnTransport> {
+    let start = Instant::now();
+    let mut set = tokio::task::JoinSet::new();
+
+    tracing::info!(
+        target_count = targets.len(),
+        interval_ms = probe_interval.as_millis(),
+        timeout_s = timeout.as_secs(),
+        "birthday: dialer starting spray"
+    );
+
+    // Spray connects with rate limiting
+    for (idx, &target) in targets.iter().enumerate() {
+        if start.elapsed() >= timeout {
+            break;
+        }
+
+        let ep = endpoint.clone();
+        let sni = call_sni.to_string();
+        let client_cfg = wzp_transport::client_config();
+        set.spawn(async move {
+            let result = wzp_transport::connect(&ep, target, &sni, client_cfg).await;
+            (idx, target, result)
+        });
+
+        // Rate limit — don't blast the NAT
+        if idx < targets.len() - 1 {
+            tokio::time::sleep(probe_interval).await;
+        }
+    }
+
+    tracing::info!(
+        spawned = set.len(),
+        elapsed_ms = start.elapsed().as_millis(),
+        "birthday: all probes spawned, waiting for first success"
+    );
+
+    // Wait for first success or all failures
+    let deadline = start + timeout;
+    while let Some(join_res) = tokio::select! {
+        r = set.join_next() => r,
+        _ = tokio::time::sleep_until(tokio::time::Instant::from_std(deadline)) => None,
+    } {
+        match join_res {
+            Ok((idx, target, Ok(conn))) => {
+                tracing::info!(
+                    idx,
+                    %target,
+                    remote = %conn.remote_address(),
+                    elapsed_ms = start.elapsed().as_millis(),
+                    "birthday: HIT! QUIC handshake succeeded"
+                );
+                set.abort_all();
+                return Some(wzp_transport::QuinnTransport::new(conn));
+            }
+            Ok((idx, target, Err(e))) => {
+                tracing::debug!(
+                    idx,
+                    %target,
+                    error = %e,
+                    "birthday: probe failed"
+                );
+            }
+            Err(_) => {}
+        }
+    }
+
+    tracing::info!(
+        elapsed_ms = start.elapsed().as_millis(),
+        "birthday: all probes failed or timed out"
+    );
+    None
+}
+
+// ── Tests ──────────────────────────────────────────────────────────
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn generate_targets_known_ports_first() {
+        let ip = Ipv4Addr::new(203, 0, 113, 5);
+        let known = vec![10000, 10001, 10002];
+        let targets = generate_dialer_targets(ip, &known, 10);
+
+        // Known ports should be first
+        assert_eq!(targets[0].port(), 10000);
+        assert_eq!(targets[1].port(), 10001);
+        assert_eq!(targets[2].port(), 10002);
+        // Rest are random
+        assert!(targets.len() <= 10);
+        // All target the right IP
+        assert!(targets.iter().all(|a| a.ip() == std::net::IpAddr::V4(ip)));
+    }
+
+    #[test]
+    fn generate_targets_no_known_all_random() {
+        let ip = Ipv4Addr::new(10, 0, 0, 1);
+        let targets = generate_dialer_targets(ip, &[], 50);
+        assert!(!targets.is_empty());
+        assert!(targets.len() <= 50);
+        // All ports in valid range
+        assert!(targets.iter().all(|a| a.port() >= 1024));
+    }
+
+    #[test]
+    fn generate_targets_more_known_than_total() {
+        let ip = Ipv4Addr::new(10, 0, 0, 1);
+        let known: Vec<u16> = (10000..10100).collect();
+        let targets = generate_dialer_targets(ip, &known, 50);
+        // All 100 known ports included even though total=50
+        assert_eq!(targets.len(), 100);
+    }
+
+    #[test]
+    fn generate_targets_dedup() {
+        let ip = Ipv4Addr::new(10, 0, 0, 1);
+        let targets = generate_dialer_targets(ip, &[], 100);
+        // No duplicates
+        let mut sorted = targets.clone();
+        sorted.sort();
+        sorted.dedup();
+        assert_eq!(sorted.len(), targets.len());
+    }
+
+    #[test]
+    fn default_config() {
+        let cfg = BirthdayConfig::default();
+        assert_eq!(cfg.acceptor_ports, 32);
+        assert_eq!(cfg.dialer_probes, 128);
+        assert!(cfg.timeout.as_secs() > 0);
+    }
+
+    #[test]
+    fn acceptor_ports_serializes() {
+        let result = AcceptorPorts {
+            external_ip: Some(Ipv4Addr::new(203, 0, 113, 5)),
+            ports: vec![PortMapping { local_port: 12345, external_port: 54321 }],
+            attempted: 32,
+            succeeded: 1,
+        };
+        let json = serde_json::to_string(&result).unwrap();
+        assert!(json.contains("54321"));
+        assert!(json.contains("203.0.113.5"));
+    }
+}

crates/wzp-client/src/call.rs

@@ -7,14 +7,15 @@ use std::time::{Duration, Instant};
 use bytes::Bytes;
 use tracing::{debug, info, warn};
 
-use wzp_codec::{AutoGainControl, ComfortNoise, EchoCanceller, NoiseSupressor, SilenceDetector};
+use wzp_codec::dred_ffi::{DredDecoderHandle, DredState};
+use wzp_codec::{
+    AdaptiveDecoder, AutoGainControl, ComfortNoise, EchoCanceller, NoiseSupressor, SilenceDetector,
+};
 use wzp_fec::{RaptorQFecDecoder, RaptorQFecEncoder};
 use wzp_proto::jitter::{JitterBuffer, PlayoutResult};
 use wzp_proto::packet::{MediaHeader, MediaPacket, MiniFrameContext};
 use wzp_proto::quality::AdaptiveQualityController;
-use wzp_proto::traits::{
-    AudioDecoder, AudioEncoder, FecDecoder, FecEncoder,
-};
+use wzp_proto::traits::{AudioDecoder, AudioEncoder, FecDecoder, FecEncoder};
 use wzp_proto::packet::QualityReport;
 use wzp_proto::{CodecId, QualityProfile};
 
@@ -42,6 +43,9 @@ pub struct CallConfig {
     /// When enabled, only every 50th frame carries a full 12-byte MediaHeader;
     /// intermediate frames use a compact 4-byte MiniHeader.
     pub mini_frames_enabled: bool,
+    /// AEC far-end delay compensation in milliseconds (default: 40).
+    /// Compensates for the round-trip audio latency from playout to mic capture.
+    pub aec_delay_ms: u32,
     /// Enable adaptive jitter buffer (default: true).
     ///
     /// When true, the jitter buffer target depth is automatically adjusted
@@ -63,6 +67,7 @@ impl Default for CallConfig {
             noise_suppression: true,
             mini_frames_enabled: true,
             adaptive_jitter: true,
+            aec_delay_ms: 40,
         }
     }
 }
@@ -229,6 +234,8 @@ pub struct CallEncoder {
     mini_frames_enabled: bool,
     /// Frames encoded since the last full header was emitted.
     frames_since_full: u32,
+    /// Pending quality report to attach to the next source packet.
+    pending_quality_report: Option<QualityReport>,
 }
 
 impl CallEncoder {
@@ -241,7 +248,7 @@ impl CallEncoder {
             block_id: 0,
             frame_in_block: 0,
             timestamp_ms: 0,
-            aec: EchoCanceller::new(48000, 100), // 100 ms echo tail
+            aec: EchoCanceller::with_delay(48000, 60, config.aec_delay_ms),
             agc: AutoGainControl::new(),
             silence_detector: SilenceDetector::new(
                 config.silence_threshold_rms,
@@ -259,6 +266,7 @@ impl CallEncoder {
             mini_context: MiniFrameContext::default(),
             mini_frames_enabled: config.mini_frames_enabled,
             frames_since_full: 0,
+            pending_quality_report: None,
         }
     }
 
@@ -340,23 +348,39 @@ impl CallEncoder {
         let enc_len = self.audio_enc.encode(pcm, &mut encoded)?;
         encoded.truncate(enc_len);
 
+        // Phase 2: Opus tiers bypass RaptorQ entirely (DRED handles loss
+        // recovery at the codec layer). Codec2 tiers keep RaptorQ unchanged.
+        // On Opus packets, zero the FEC header fields so old receivers
+        // can cleanly identify "no RaptorQ block to assemble" and new
+        // receivers can short-circuit their FEC ingest path.
+        let is_opus = self.profile.codec.is_opus();
+        let (fec_block, fec_symbol, fec_ratio_encoded) = if is_opus {
+            (0u8, 0u8, 0u8)
+        } else {
+            (
+                self.block_id,
+                self.frame_in_block,
+                MediaHeader::encode_fec_ratio(self.profile.fec_ratio),
+            )
+        };
+
         // Build source media packet
         let source_pkt = MediaPacket {
             header: MediaHeader {
                 version: 0,
                 is_repair: false,
                 codec_id: self.profile.codec,
-                has_quality_report: false,
-                fec_ratio_encoded: MediaHeader::encode_fec_ratio(self.profile.fec_ratio),
+                has_quality_report: self.pending_quality_report.is_some(),
+                fec_ratio_encoded,
                 seq: self.seq,
                 timestamp: self.timestamp_ms,
-                fec_block: self.block_id,
-                fec_symbol: self.frame_in_block,
+                fec_block,
+                fec_symbol,
                 reserved: 0,
                 csrc_count: 0,
             },
             payload: Bytes::from(encoded.clone()),
-            quality_report: None,
+            quality_report: self.pending_quality_report.take(),
         };
 
         self.seq = self.seq.wrapping_add(1);
@@ -366,39 +390,42 @@ impl CallEncoder {
 
         let mut output = vec![source_pkt];
 
-        // Add to FEC encoder
-        self.fec_enc.add_source_symbol(&encoded)?;
-        self.frame_in_block += 1;
+        // Codec2-only: feed RaptorQ and generate repair packets when the
+        // block is full. Opus tiers skip this entire block — DRED (active
+        // in Phase 1) provides codec-layer loss recovery.
+        if !is_opus {
+            self.fec_enc.add_source_symbol(&encoded)?;
+            self.frame_in_block += 1;
 
-        // If block is full, generate repair and finalize
-        if self.frame_in_block >= self.profile.frames_per_block {
-            if let Ok(repairs) = self.fec_enc.generate_repair(self.profile.fec_ratio) {
-                for (sym_idx, repair_data) in repairs {
-                    output.push(MediaPacket {
-                        header: MediaHeader {
-                            version: 0,
-                            is_repair: true,
-                            codec_id: self.profile.codec,
-                            has_quality_report: false,
-                            fec_ratio_encoded: MediaHeader::encode_fec_ratio(
-                                self.profile.fec_ratio,
-                            ),
-                            seq: self.seq,
-                            timestamp: self.timestamp_ms,
-                            fec_block: self.block_id,
-                            fec_symbol: sym_idx,
-                            reserved: 0,
-                            csrc_count: 0,
-                        },
-                        payload: Bytes::from(repair_data),
-                        quality_report: None,
-                    });
-                    self.seq = self.seq.wrapping_add(1);
+            if self.frame_in_block >= self.profile.frames_per_block {
+                if let Ok(repairs) = self.fec_enc.generate_repair(self.profile.fec_ratio) {
+                    for (sym_idx, repair_data) in repairs {
+                        output.push(MediaPacket {
+                            header: MediaHeader {
+                                version: 0,
+                                is_repair: true,
+                                codec_id: self.profile.codec,
+                                has_quality_report: false,
+                                fec_ratio_encoded: MediaHeader::encode_fec_ratio(
+                                    self.profile.fec_ratio,
+                                ),
+                                seq: self.seq,
+                                timestamp: self.timestamp_ms,
+                                fec_block: self.block_id,
+                                fec_symbol: sym_idx,
+                                reserved: 0,
+                                csrc_count: 0,
+                            },
+                            payload: Bytes::from(repair_data),
+                            quality_report: None,
+                        });
+                        self.seq = self.seq.wrapping_add(1);
+                    }
                 }
+                let _ = self.fec_enc.finalize_block();
+                self.block_id = self.block_id.wrapping_add(1);
+                self.frame_in_block = 0;
             }
-            let _ = self.fec_enc.finalize_block();
-            self.block_id = self.block_id.wrapping_add(1);
-            self.frame_in_block = 0;
         }
 
         Ok(output)
@@ -421,6 +448,22 @@ impl CallEncoder {
         self.aec.feed_farend(farend);
     }
 
+    /// Apply DRED tuning output to the encoder.
+    ///
+    /// Called by the send loop after `DredTuner::update()` returns `Some`.
+    /// No-op when the active codec is Codec2 (DRED is Opus-only).
+    pub fn apply_dred_tuning(&mut self, tuning: wzp_proto::DredTuning) {
+        self.audio_enc.set_dred_duration(tuning.dred_frames);
+        self.audio_enc.set_expected_loss(tuning.expected_loss_pct);
+    }
+
+    /// Queue a quality report for attachment to the next source packet.
+    /// Used by the send task to embed locally-observed path quality so
+    /// the peer can drive adaptive quality switching.
+    pub fn set_pending_quality_report(&mut self, report: QualityReport) {
+        self.pending_quality_report = Some(report);
+    }
+
     /// Enable or disable acoustic echo cancellation.
     pub fn set_aec_enabled(&mut self, enabled: bool) {
         self.aec.set_enabled(enabled);
@@ -434,9 +477,12 @@ impl CallEncoder {
 
 /// Manages the recv/decode side of a call.
 pub struct CallDecoder {
-    /// Audio decoder.
-    audio_dec: Box<dyn AudioDecoder>,
-    /// FEC decoder.
+    /// Audio decoder. Concrete `AdaptiveDecoder` (not `Box<dyn AudioDecoder>`)
+    /// because Phase 3b calls the inherent `reconstruct_from_dred` method,
+    /// which cannot live on the `AudioDecoder` trait without dragging libopus
+    /// types into `wzp-proto`.
+    audio_dec: AdaptiveDecoder,
+    /// FEC decoder (Codec2 tiers only; Opus bypasses RaptorQ per Phase 2).
     fec_dec: RaptorQFecDecoder,
     /// Jitter buffer.
     jitter: JitterBuffer,
@@ -450,6 +496,24 @@ pub struct CallDecoder {
     last_was_cn: bool,
     /// Mini-frame decompression context (tracks last full header baseline).
     mini_context: MiniFrameContext,
+    // ─── Phase 3b: DRED reconstruction state ──────────────────────────────
+    /// DRED side-channel parser (a separate libopus object from the decoder).
+    dred_decoder: DredDecoderHandle,
+    /// Scratch buffer used by `dred_decoder.parse_into` on every arriving
+    /// Opus packet. Reused across calls to avoid 10 KB alloc churn per packet.
+    dred_parse_scratch: DredState,
+    /// Cached "most recently parsed valid" DRED state, swapped with
+    /// `dred_parse_scratch` on successful parse. Used by `decode_next` when
+    /// the jitter buffer reports a gap.
+    last_good_dred: DredState,
+    /// Sequence number of the packet that produced `last_good_dred`. `None`
+    /// if no packet has yielded DRED state yet (cold start or legacy sender).
+    last_good_dred_seq: Option<u16>,
+    /// Phase 4 telemetry counter: gaps recovered via DRED reconstruction.
+    pub dred_reconstructions: u64,
+    /// Phase 4 telemetry counter: gaps filled via classical Opus PLC
+    /// (because no DRED state covered the gap, or the active codec is Codec2).
+    pub classical_plc_invocations: u64,
 }
 
 impl CallDecoder {
@@ -459,8 +523,19 @@ impl CallDecoder {
         } else {
             JitterBuffer::new(config.jitter_target, config.jitter_max, config.jitter_min)
         };
+        // Phase 3b: build the DRED parser + state buffers. These allocate
+        // libopus state (~10 KB each) once per call, not per packet — the
+        // scratch and last-good buffers are reused via std::mem::swap on
+        // every successful parse.
+        let dred_decoder =
+            DredDecoderHandle::new().expect("opus_dred_decoder_create failed at call setup");
+        let dred_parse_scratch =
+            DredState::new().expect("opus_dred_alloc failed at call setup (scratch)");
+        let last_good_dred =
+            DredState::new().expect("opus_dred_alloc failed at call setup (good state)");
         Self {
-            audio_dec: wzp_codec::create_decoder(config.profile),
+            audio_dec: AdaptiveDecoder::new(config.profile)
+                .expect("failed to create adaptive decoder"),
             fec_dec: wzp_fec::create_decoder(&config.profile),
             jitter,
             quality: AdaptiveQualityController::new(),
@@ -468,6 +543,12 @@ impl CallDecoder {
             comfort_noise: ComfortNoise::new(50),
             last_was_cn: false,
             mini_context: MiniFrameContext::default(),
+            dred_decoder,
+            dred_parse_scratch,
+            last_good_dred,
+            last_good_dred_seq: None,
+            dred_reconstructions: 0,
+            classical_plc_invocations: 0,
         }
     }
 
@@ -482,20 +563,105 @@ impl CallDecoder {
 
     /// Feed a received media packet into the decode pipeline.
     pub fn ingest(&mut self, packet: MediaPacket) {
-        // Feed to FEC decoder
-        let _ = self.fec_dec.add_symbol(
-            packet.header.fec_block,
-            packet.header.fec_symbol,
-            packet.header.is_repair,
-            &packet.payload,
-        );
+        // Phase 2: Opus packets bypass RaptorQ. Codec2 packets still feed
+        // the FEC decoder for recovery. This also cleanly drops any stray
+        // Opus repair packets from an old sender (we don't push repair
+        // packets to the jitter buffer either, so they're effectively
+        // ignored — a graceful mixed-version degradation).
+        if !packet.header.codec_id.is_opus() {
+            let _ = self.fec_dec.add_symbol(
+                packet.header.fec_block,
+                packet.header.fec_symbol,
+                packet.header.is_repair,
+                &packet.payload,
+            );
+        }
 
-        // If not a repair packet, also feed directly to jitter buffer
+        // Phase 3b: Opus source packets carry DRED side-channel data in
+        // libopus 1.5. Parse it into the scratch state and, on success,
+        // swap with the cached `last_good_dred` so later gap reconstruction
+        // has fresh neural redundancy to draw from. Parsing happens before
+        // the jitter push because the jitter buffer consumes the packet.
+        if packet.header.codec_id.is_opus() && !packet.header.is_repair {
+            match self
+                .dred_decoder
+                .parse_into(&mut self.dred_parse_scratch, &packet.payload)
+            {
+                Ok(available) if available > 0 => {
+                    // Swap the freshly parsed state into `last_good_dred`.
+                    // The old good state (now in scratch) is about to be
+                    // overwritten on the next parse — its contents are
+                    // not needed after this swap.
+                    std::mem::swap(&mut self.dred_parse_scratch, &mut self.last_good_dred);
+                    self.last_good_dred_seq = Some(packet.header.seq);
+                }
+                Ok(_) => {
+                    // Packet had no DRED data (return 0). Leave the cached
+                    // state untouched — it may still cover upcoming gaps
+                    // from a warm-up period where the encoder was producing
+                    // DRED bytes. The scratch buffer was potentially written
+                    // but its `samples_available` is 0 so it's harmless.
+                }
+                Err(e) => {
+                    debug!("DRED parse error (ignored): {e}");
+                }
+            }
+        }
+
+        // Source packets (Opus or Codec2) go to the jitter buffer for decode.
+        // Repair packets never reach the jitter buffer; for Codec2 they're
+        // used by the FEC decoder above, for Opus they're dropped here.
         if !packet.header.is_repair {
             self.jitter.push(packet);
         }
     }
 
+    /// Switch the decoder to match an incoming packet's codec if it differs
+    /// from the current profile. This enables cross-codec interop (e.g. one
+    /// client sends Opus, the other sends Codec2).
+    fn switch_decoder_if_needed(&mut self, incoming_codec: CodecId) {
+        if incoming_codec == self.profile.codec || incoming_codec == CodecId::ComfortNoise {
+            return;
+        }
+        let new_profile = Self::profile_for_codec(incoming_codec);
+        info!(
+            from = ?self.profile.codec,
+            to = ?incoming_codec,
+            "decoder switching codec to match incoming packet"
+        );
+        if let Err(e) = self.audio_dec.set_profile(new_profile) {
+            warn!("failed to switch decoder profile: {e}");
+            return;
+        }
+        self.fec_dec = wzp_fec::create_decoder(&new_profile);
+        self.profile = new_profile;
+    }
+
+    /// Map a `CodecId` to a reasonable `QualityProfile` for decoding.
+    fn profile_for_codec(codec: CodecId) -> QualityProfile {
+        match codec {
+            CodecId::Opus24k => QualityProfile::GOOD,
+            CodecId::Opus16k => QualityProfile {
+                codec: CodecId::Opus16k,
+                fec_ratio: 0.3,
+                frame_duration_ms: 20,
+                frames_per_block: 5,
+            },
+            CodecId::Opus6k => QualityProfile::DEGRADED,
+            CodecId::Opus32k => QualityProfile::STUDIO_32K,
+            CodecId::Opus48k => QualityProfile::STUDIO_48K,
+            CodecId::Opus64k => QualityProfile::STUDIO_64K,
+            CodecId::Codec2_3200 => QualityProfile {
+                codec: CodecId::Codec2_3200,
+                fec_ratio: 0.5,
+                frame_duration_ms: 20,
+                frames_per_block: 5,
+            },
+            CodecId::Codec2_1200 => QualityProfile::CATASTROPHIC,
+            CodecId::ComfortNoise => QualityProfile::GOOD,
+        }
+    }
+
     /// Decode the next audio frame from the jitter buffer.
     ///
     /// Returns PCM samples (48kHz mono) or None if not ready.
@@ -510,6 +676,9 @@ impl CallDecoder {
                     return Some(pcm.len());
                 }
 
+                // Auto-switch decoder if incoming codec differs from current.
+                self.switch_decoder_if_needed(pkt.header.codec_id);
+
                 self.last_was_cn = false;
                 let result = match self.audio_dec.decode(&pkt.payload, pcm) {
                     Ok(n) => Some(n),
@@ -524,19 +693,72 @@ impl CallDecoder {
                 result
             }
             PlayoutResult::Missing { seq } => {
-                // Only generate PLC if there are still packets buffered ahead.
+                // Only attempt recovery if there are still packets buffered ahead.
                 // Otherwise we've drained everything — return None to stop.
-                if self.jitter.depth() > 0 {
-                    debug!(seq, "packet loss, generating PLC");
-                    let result = self.audio_dec.decode_lost(pcm).ok();
-                    if result.is_some() {
-                        self.jitter.record_decode();
-                    }
-                    result
-                } else {
+                if self.jitter.depth() == 0 {
                     self.jitter.record_underrun();
-                    None
+                    return None;
                 }
+
+                // Phase 3b: try DRED reconstruction first. If we have a
+                // recent DRED state from a packet whose seq > missing seq,
+                // and the seq delta (in samples) fits within the state's
+                // available window, libopus can synthesize a plausible
+                // replacement for the lost frame. Fall back to classical
+                // PLC when no state covers the gap, when the active codec
+                // is Codec2, or when the reconstruction itself errors.
+                if self.profile.codec.is_opus() {
+                    if let Some(last_seq) = self.last_good_dred_seq {
+                        // How many frames ahead of the missing seq is the
+                        // last-good packet? Use wrapping arithmetic for the
+                        // u16 seq space.
+                        let seq_delta = last_seq.wrapping_sub(seq);
+                        // Reject stale or backward state. u16 wraparound
+                        // would make a "seq went backward" delta very large;
+                        // cap at a sane forward-looking window.
+                        const MAX_SEQ_DELTA: u16 = 128;
+                        if seq_delta > 0 && seq_delta <= MAX_SEQ_DELTA {
+                            let frame_samples =
+                                (48_000 * self.profile.frame_duration_ms as i32) / 1000;
+                            let offset_samples = seq_delta as i32 * frame_samples;
+                            let available = self.last_good_dred.samples_available();
+                            if offset_samples > 0 && offset_samples <= available {
+                                match self.audio_dec.reconstruct_from_dred(
+                                    &self.last_good_dred,
+                                    offset_samples,
+                                    pcm,
+                                ) {
+                                    Ok(n) => {
+                                        self.dred_reconstructions += 1;
+                                        self.jitter.record_decode();
+                                        debug!(
+                                            seq,
+                                            last_seq,
+                                            offset_samples,
+                                            available,
+                                            "DRED reconstruction for gap"
+                                        );
+                                        return Some(n);
+                                    }
+                                    Err(e) => {
+                                        // Reconstruction failed — fall
+                                        // through to classical PLC below.
+                                        debug!(seq, "DRED reconstruct error: {e}");
+                                    }
+                                }
+                            }
+                        }
+                    }
+                }
+
+                // Classical PLC fallback (also the Codec2 path).
+                debug!(seq, "packet loss, generating classical PLC");
+                self.classical_plc_invocations += 1;
+                let result = self.audio_dec.decode_lost(pcm).ok();
+                if result.is_some() {
+                    self.jitter.record_decode();
+                }
+                result
             }
             PlayoutResult::NotReady => {
                 self.jitter.record_underrun();
@@ -559,6 +781,19 @@ impl CallDecoder {
     pub fn reset_stats(&mut self) {
         self.jitter.reset_stats();
     }
+
+    /// Phase 3b introspection: sequence number of the most recently parsed
+    /// valid DRED state, or `None` if no Opus packet has yielded DRED data
+    /// yet. Used by tests to debug reconstruction eligibility.
+    pub fn last_good_dred_seq(&self) -> Option<u16> {
+        self.last_good_dred_seq
+    }
+
+    /// Phase 3b introspection: samples of audio history currently available
+    /// in the cached DRED state.
+    pub fn last_good_dred_samples_available(&self) -> i32 {
+        self.last_good_dred.samples_available()
+    }
 }
 
 /// Periodic telemetry logger for jitter buffer statistics.
@@ -620,18 +855,83 @@ mod tests {
         assert!(!packets[0].header.is_repair);
     }
 
+    /// Phase 2: Opus packets have zero FEC header fields — no block, no
+    /// symbol index, no repair ratio. The RaptorQ layer is bypassed
+    /// entirely on the Opus tiers.
     #[test]
-    fn encoder_generates_repair_on_full_block() {
+    fn opus_source_packets_have_zero_fec_header_fields() {
         let config = CallConfig {
-            profile: QualityProfile::GOOD, // 5 frames/block
+            profile: QualityProfile::GOOD, // Opus 24k
+            suppression_enabled: false,    // skip silence gate for this test
             ..Default::default()
         };
         let mut enc = CallEncoder::new(&config);
-        let pcm = vec![0i16; 960];
+        // Non-silent sine wave so silence detection doesn't suppress us
+        // even with suppression_enabled=false (belt and braces).
+        let pcm: Vec<i16> = (0..960)
+            .map(|i| ((i as f32 * 0.1).sin() * 10_000.0) as i16)
+            .collect();
+        let packets = enc.encode_frame(&pcm).unwrap();
+        assert_eq!(packets.len(), 1, "Opus must emit exactly 1 source packet");
+        let hdr = &packets[0].header;
+        assert!(hdr.codec_id.is_opus());
+        assert!(!hdr.is_repair);
+        assert_eq!(hdr.fec_block, 0, "Opus fec_block must be 0");
+        assert_eq!(hdr.fec_symbol, 0, "Opus fec_symbol must be 0");
+        assert_eq!(hdr.fec_ratio_encoded, 0, "Opus fec_ratio_encoded must be 0");
+    }
 
-        let mut total_packets = 0;
-        let mut repair_count = 0;
-        for _ in 0..5 {
+    /// Phase 2: Opus never emits repair packets, regardless of how many
+    /// source frames are fed in. DRED (Phase 1) provides loss recovery at
+    /// the codec layer; RaptorQ is disabled on Opus tiers.
+    #[test]
+    fn opus_encoder_never_emits_repair_packets() {
+        let config = CallConfig {
+            profile: QualityProfile::GOOD, // 5 frames/block in the Codec2 sense
+            suppression_enabled: false,
+            ..Default::default()
+        };
+        let mut enc = CallEncoder::new(&config);
+        let pcm: Vec<i16> = (0..960)
+            .map(|i| ((i as f32 * 0.1).sin() * 10_000.0) as i16)
+            .collect();
+
+        // Encode well beyond a block boundary to prove no repair ever comes out.
+        let mut total_packets = 0usize;
+        let mut repair_count = 0usize;
+        for _ in 0..20 {
+            let packets = enc.encode_frame(&pcm).unwrap();
+            total_packets += packets.len();
+            repair_count += packets.iter().filter(|p| p.header.is_repair).count();
+        }
+        assert_eq!(repair_count, 0, "Opus must emit zero repair packets");
+        assert_eq!(
+            total_packets, 20,
+            "20 source frames → 20 source packets (1:1, no RaptorQ expansion)"
+        );
+    }
+
+    /// Phase 2: Codec2 still emits repair packets with RaptorQ ratio unchanged.
+    /// DRED is libopus-only and does not apply here, so RaptorQ is still the
+    /// primary loss-recovery mechanism on Codec2 tiers.
+    #[test]
+    fn codec2_encoder_generates_repair_on_full_block() {
+        let config = CallConfig {
+            profile: QualityProfile::CATASTROPHIC, // Codec2 1200, 8 frames/block, ratio 1.0
+            suppression_enabled: false,
+            ..Default::default()
+        };
+        let mut enc = CallEncoder::new(&config);
+        // Codec2 takes 48 kHz samples and downsamples internally.
+        // CATASTROPHIC uses 40 ms frames → 1920 samples.
+        let pcm: Vec<i16> = (0..1920)
+            .map(|i| ((i as f32 * 0.1).sin() * 10_000.0) as i16)
+            .collect();
+
+        let mut total_packets = 0usize;
+        let mut repair_count = 0usize;
+        // Run long enough to cross the 8-frame block boundary and see repairs.
+        for _ in 0..16 {
             let packets = enc.encode_frame(&pcm).unwrap();
             for p in &packets {
                 if p.header.is_repair {
@@ -640,8 +940,10 @@ mod tests {
             }
             total_packets += packets.len();
         }
-        assert!(repair_count > 0, "should have repair packets after full block");
-        assert!(total_packets > 5, "total {total_packets} should exceed 5 source");
+        assert!(
+            repair_count > 0,
+            "Codec2 must still emit repair packets (got {repair_count} repairs, {total_packets} total)"
+        );
     }
 
     #[test]
@@ -672,6 +974,219 @@ mod tests {
         assert!(dec.decode_next(&mut pcm).is_none());
     }
 
+    // ─── Phase 3b — DRED reconstruction on packet loss ────────────────────
+
+    /// Helper: create a CallEncoder/CallDecoder pair with the given profile
+    /// and silence suppression disabled so silence-detection doesn't drop
+    /// our synthetic test frames.
+    fn encoder_decoder_pair(profile: QualityProfile) -> (CallEncoder, CallDecoder) {
+        let config = CallConfig {
+            profile,
+            suppression_enabled: false,
+            // Small jitter buffer so decode_next drains quickly in tests.
+            jitter_min: 2,
+            jitter_target: 3,
+            jitter_max: 20,
+            adaptive_jitter: false,
+            ..Default::default()
+        };
+        (CallEncoder::new(&config), CallDecoder::new(&config))
+    }
+
+    /// Helper: generate a non-silent 20 ms frame of 300 Hz sine at the
+    /// given sample offset so consecutive frames form a continuous tone.
+    fn voice_frame_20ms(sample_offset: usize) -> Vec<i16> {
+        (0..960)
+            .map(|i| {
+                let t = (sample_offset + i) as f64 / 48_000.0;
+                (8000.0 * (2.0 * std::f64::consts::PI * 300.0 * t).sin()) as i16
+            })
+            .collect()
+    }
+
+    /// Phase 3b probe: sweep packet_loss_perc values to find the minimum
+    /// that produces a samples_available ≥ 960 (enough to reconstruct a
+    /// single 20 ms Opus frame). This guides the production loss floor.
+    #[test]
+    #[ignore] // diagnostic only — run with `cargo test ... -- --ignored --nocapture`
+    fn probe_dred_samples_available_by_loss_floor() {
+        use wzp_codec::opus_enc::OpusEncoder;
+        use wzp_proto::traits::AudioEncoder;
+
+        for loss_pct in [5u8, 10, 15, 20, 25, 40, 60, 80].iter().copied() {
+            let mut enc = OpusEncoder::new(QualityProfile::GOOD).unwrap();
+            enc.set_expected_loss(loss_pct);
+            let (_drop_enc, mut dec) = encoder_decoder_pair(QualityProfile::GOOD);
+
+            for i in 0..60u16 {
+                let pcm = voice_frame_20ms(i as usize * 960);
+                let mut encoded = vec![0u8; 512];
+                let n = enc.encode(&pcm, &mut encoded).unwrap();
+                encoded.truncate(n);
+                let pkt = MediaPacket {
+                    header: MediaHeader {
+                        version: 0,
+                        is_repair: false,
+                        codec_id: CodecId::Opus24k,
+                        has_quality_report: false,
+                        fec_ratio_encoded: 0,
+                        seq: i,
+                        timestamp: (i as u32) * 20,
+                        fec_block: 0,
+                        fec_symbol: 0,
+                        reserved: 0,
+                        csrc_count: 0,
+                    },
+                    payload: Bytes::from(encoded),
+                    quality_report: None,
+                };
+                dec.ingest(pkt);
+            }
+            eprintln!(
+                "[phase3b probe] loss_pct={loss_pct} samples_available={}",
+                dec.last_good_dred_samples_available()
+            );
+        }
+    }
+
+    /// Phase 3b: simulated single-packet loss on an Opus call triggers a
+    /// DRED reconstruction rather than a classical PLC fill. Runs the full
+    /// encode → ingest → decode_next pipeline.
+    #[test]
+    fn opus_single_packet_loss_is_recovered_via_dred() {
+        let (mut enc, mut dec) = encoder_decoder_pair(QualityProfile::GOOD);
+
+        // Warm-up: encode and ingest 60 frames (1.2 s) so the DRED emitter
+        // has had time to fill its 200 ms window and at least one
+        // successful DRED parse has happened on the decoder side.
+        let warmup_frames = 60;
+        for i in 0..warmup_frames {
+            let pcm = voice_frame_20ms(i * 960);
+            let packets = enc.encode_frame(&pcm).unwrap();
+            for pkt in packets {
+                dec.ingest(pkt);
+            }
+        }
+
+        // Drain the warm-up frames through the decoder to advance the
+        // jitter buffer cursor past them.
+        let mut out = vec![0i16; 960];
+        while dec.decode_next(&mut out).is_some() {}
+
+        // Encode the next three frames but skip ingesting the middle one.
+        let base_offset = warmup_frames * 960;
+        let pcm_a = voice_frame_20ms(base_offset);
+        let pcm_b = voice_frame_20ms(base_offset + 960);
+        let pcm_c = voice_frame_20ms(base_offset + 1920);
+
+        let pkts_a = enc.encode_frame(&pcm_a).unwrap();
+        let pkts_b = enc.encode_frame(&pcm_b).unwrap(); // DROP THIS ONE
+        let pkts_c = enc.encode_frame(&pcm_c).unwrap();
+
+        for pkt in pkts_a {
+            dec.ingest(pkt);
+        }
+        // Skip pkts_b entirely — this is the "packet loss".
+        drop(pkts_b);
+        for pkt in pkts_c {
+            dec.ingest(pkt);
+        }
+
+        // Drain again. Somewhere in here decode_next will hit Missing()
+        // for the dropped packet and attempt DRED reconstruction.
+        let baseline_dred = dec.dred_reconstructions;
+        let baseline_plc = dec.classical_plc_invocations;
+        eprintln!(
+            "[phase3b probe] pre-drain: last_good_seq={:?} samples_available={}",
+            dec.last_good_dred_seq(),
+            dec.last_good_dred_samples_available()
+        );
+        while dec.decode_next(&mut out).is_some() {}
+
+        let dred_delta = dec.dred_reconstructions - baseline_dred;
+        let plc_delta = dec.classical_plc_invocations - baseline_plc;
+        eprintln!(
+            "[phase3b probe] post-drain: dred_delta={dred_delta} plc_delta={plc_delta}"
+        );
+        assert!(
+            dred_delta >= 1,
+            "expected ≥1 DRED reconstruction on single-packet loss, \
+             got dred_delta={dred_delta} plc_delta={plc_delta}"
+        );
+    }
+
+    /// Phase 3b: lossless stream never triggers DRED reconstruction or PLC.
+    /// Baseline behavior — verifies the Missing() branch is not spuriously taken.
+    #[test]
+    fn opus_lossless_ingest_never_triggers_dred_or_plc() {
+        let (mut enc, mut dec) = encoder_decoder_pair(QualityProfile::GOOD);
+
+        // Encode + ingest 40 frames with no drops.
+        for i in 0..40 {
+            let pcm = voice_frame_20ms(i * 960);
+            let packets = enc.encode_frame(&pcm).unwrap();
+            for pkt in packets {
+                dec.ingest(pkt);
+            }
+        }
+
+        let mut out = vec![0i16; 960];
+        while dec.decode_next(&mut out).is_some() {}
+
+        assert_eq!(
+            dec.dred_reconstructions, 0,
+            "lossless stream should not reconstruct"
+        );
+        assert_eq!(
+            dec.classical_plc_invocations, 0,
+            "lossless stream should not PLC"
+        );
+    }
+
+    /// Phase 3b: Codec2 calls fall through to classical PLC on loss.
+    /// DRED is libopus-only, so even if the decoder's DRED state were
+    /// populated (it won't be — Codec2 packets don't carry DRED bytes),
+    /// `reconstruct_from_dred` rejects Codec2 at the AdaptiveDecoder
+    /// level. This test guards the Codec2 side of the protection split.
+    #[test]
+    fn codec2_loss_falls_through_to_classical_plc() {
+        let (mut enc, mut dec) = encoder_decoder_pair(QualityProfile::CATASTROPHIC);
+
+        // Codec2 1200 uses 40 ms frames → 1920 samples at 48 kHz (before
+        // the downsample inside the codec). Encode 20 frames (~0.8 s).
+        let make_frame = |offset: usize| -> Vec<i16> {
+            (0..1920)
+                .map(|i| {
+                    let t = (offset + i) as f64 / 48_000.0;
+                    (8000.0 * (2.0 * std::f64::consts::PI * 300.0 * t).sin()) as i16
+                })
+                .collect()
+        };
+
+        for i in 0..20 {
+            let pcm = make_frame(i * 1920);
+            let packets = enc.encode_frame(&pcm).unwrap();
+            for pkt in packets {
+                // Drop every 5th source packet to simulate loss.
+                if !pkt.header.is_repair && i % 5 == 3 {
+                    continue;
+                }
+                dec.ingest(pkt);
+            }
+        }
+
+        let mut out = vec![0i16; 1920];
+        while dec.decode_next(&mut out).is_some() {}
+
+        assert_eq!(
+            dec.dred_reconstructions, 0,
+            "Codec2 must never reconstruct via DRED"
+        );
+        // classical_plc_invocations may or may not trigger depending on
+        // whether the jitter buffer sees Missing before draining — the key
+        // assertion is that DRED is not used. PLC count is advisory.
+    }
+
     // ---- QualityAdapter tests ----
 
     /// Helper: build a QualityReport from human-readable loss% and RTT ms.
@@ -946,4 +1461,155 @@ mod tests {
             "frames_suppressed should be > 0"
         );
     }
+
+    // ---- DredTuner integration tests ----
+
+    /// End-to-end test: DredTuner reacts to simulated network degradation
+    /// and adjusts the encoder's DRED parameters via `apply_dred_tuning`.
+    #[test]
+    fn dred_tuner_adjusts_encoder_on_loss() {
+        use wzp_proto::DredTuner;
+
+        let mut enc = CallEncoder::new(&CallConfig {
+            profile: QualityProfile::GOOD,
+            suppression_enabled: false,
+            ..Default::default()
+        });
+        let mut tuner = DredTuner::new(QualityProfile::GOOD.codec);
+
+        // Baseline: good network → baseline DRED (20 frames = 200 ms).
+        let baseline = tuner.current();
+        assert_eq!(baseline.dred_frames, 20);
+
+        // Warm up the tuner — first few updates may return Some as the
+        // EWMA initializes and expected_loss settles from the initial 15%.
+        for _ in 0..10 {
+            tuner.update(0.0, 50, 5);
+        }
+        // After settling, the tuning should be at baseline.
+        assert_eq!(tuner.current().dred_frames, 20);
+
+        // Simulate network degradation: 30% loss, 300ms RTT.
+        // The tuner should increase DRED frames above baseline.
+        let tuning = tuner.update(30.0, 300, 15);
+        assert!(tuning.is_some(), "loss spike should trigger tuning change");
+        let t = tuning.unwrap();
+        assert!(
+            t.dred_frames > 20,
+            "30% loss should increase DRED above baseline 20, got {}",
+            t.dred_frames
+        );
+
+        // Apply to encoder — should not panic.
+        enc.apply_dred_tuning(t);
+
+        // Verify the encoder still works after tuning.
+        let pcm = voice_frame_20ms(0);
+        let packets = enc.encode_frame(&pcm).unwrap();
+        assert!(!packets.is_empty(), "encoder must still produce packets after DRED tuning");
+    }
+
+    /// DredTuner jitter spike triggers pre-emptive DRED boost to ceiling.
+    #[test]
+    fn dred_tuner_spike_boosts_to_ceiling() {
+        use wzp_proto::DredTuner;
+
+        let mut tuner = DredTuner::new(CodecId::Opus24k);
+
+        // Establish low-jitter baseline.
+        for _ in 0..20 {
+            tuner.update(0.0, 50, 5);
+        }
+        assert!(!tuner.spike_boost_active());
+
+        // Jitter spikes to 40ms (8x baseline of ~5ms).
+        let tuning = tuner.update(0.0, 50, 40);
+        assert!(tuner.spike_boost_active(), "jitter spike should activate boost");
+        assert!(tuning.is_some());
+        // Ceiling for Opus24k is 50 frames = 500 ms.
+        assert_eq!(
+            tuning.unwrap().dred_frames, 50,
+            "spike should push to ceiling"
+        );
+    }
+
+    /// DredTuner is a no-op for Codec2 profiles.
+    #[test]
+    fn dred_tuner_noop_for_codec2() {
+        use wzp_proto::DredTuner;
+
+        let mut tuner = DredTuner::new(CodecId::Codec2_1200);
+
+        // Even extreme conditions produce no tuning output.
+        assert!(tuner.update(50.0, 800, 100).is_none());
+        assert_eq!(tuner.current().dred_frames, 0);
+    }
+
+    /// DredTuner + CallEncoder: full cycle through profile switch.
+    #[test]
+    fn dred_tuner_handles_profile_switch() {
+        use wzp_proto::DredTuner;
+
+        let mut enc = CallEncoder::new(&CallConfig {
+            profile: QualityProfile::GOOD,
+            suppression_enabled: false,
+            ..Default::default()
+        });
+        let mut tuner = DredTuner::new(QualityProfile::GOOD.codec);
+
+        // Apply initial tuning on good network.
+        if let Some(t) = tuner.update(0.0, 50, 5) {
+            enc.apply_dred_tuning(t);
+        }
+
+        // Switch to degraded profile.
+        enc.set_profile(QualityProfile::DEGRADED).unwrap();
+        tuner.set_codec(QualityProfile::DEGRADED.codec);
+
+        // Opus6k baseline is 50 frames (500 ms), ceiling is 104 (1040 ms).
+        let baseline = tuner.current();
+        // After set_codec, the cached tuning should reflect old state;
+        // a fresh update gives the new codec's mapping.
+        let tuning = tuner.update(20.0, 200, 10);
+        assert!(tuning.is_some());
+        let t = tuning.unwrap();
+        assert!(
+            t.dred_frames >= 50,
+            "Opus6k with 20% loss should be at least baseline 50, got {}",
+            t.dred_frames
+        );
+
+        enc.apply_dred_tuning(t);
+
+        // Encode a 40ms frame (Opus6k uses 40ms frames = 1920 samples).
+        let pcm: Vec<i16> = (0..1920)
+            .map(|i| ((i as f32 * 0.1).sin() * 10_000.0) as i16)
+            .collect();
+        let packets = enc.encode_frame(&pcm).unwrap();
+        assert!(!packets.is_empty());
+    }
+
+    #[test]
+    fn encoder_attaches_quality_report() {
+        let mut enc = CallEncoder::new(&CallConfig {
+            profile: QualityProfile::GOOD,
+            suppression_enabled: false,
+            ..Default::default()
+        });
+
+        // Set a quality report
+        enc.set_pending_quality_report(QualityReport::from_path_stats(5.0, 80, 10));
+
+        // Encode a frame — should have quality_report attached
+        let pcm = voice_frame_20ms(0);
+        let packets = enc.encode_frame(&pcm).unwrap();
+        assert!(!packets.is_empty());
+        assert!(packets[0].header.has_quality_report, "first packet should have quality report");
+        assert!(packets[0].quality_report.is_some());
+
+        // Next frame should NOT have quality_report (it was consumed)
+        let packets2 = enc.encode_frame(&voice_frame_20ms(960)).unwrap();
+        assert!(!packets2[0].header.has_quality_report, "second packet should not have quality report");
+        assert!(packets2[0].quality_report.is_none());
+    }
 }

crates/wzp-client/src/cli.rs

@@ -52,6 +52,8 @@ struct CliArgs {
     signal: bool,
     /// Place a direct call to a fingerprint (requires --signal).
     call_target: Option<String>,
+    /// Run network diagnostic (STUN, port mapping, relay latencies).
+    netcheck: bool,
 }
 
 impl CliArgs {
@@ -97,6 +99,7 @@ fn parse_args() -> CliArgs {
     let mut relay_str = None;
     let mut signal = false;
     let mut call_target = None;
+    let mut netcheck = false;
 
     let mut i = 1;
     while i < args.len() {
@@ -182,6 +185,7 @@ fn parse_args() -> CliArgs {
                 );
             }
             "--sweep" => sweep = true,
+            "--netcheck" => { netcheck = true; }
             "--version-check" => { version_check = true; }
             "--help" | "-h" => {
                 eprintln!("Usage: wzp-client [options] [relay-addr]");
@@ -238,6 +242,7 @@ fn parse_args() -> CliArgs {
         version_check,
         signal,
         call_target,
+        netcheck,
     }
 }
 
@@ -256,6 +261,23 @@ async fn main() -> anyhow::Result<()> {
         return Ok(());
     }
 
+    // --netcheck: run network diagnostic and exit
+    if cli.netcheck {
+        let config = wzp_client::netcheck::NetcheckConfig {
+            stun_config: wzp_client::stun::StunConfig::default(),
+            relays: vec![
+                ("relay".into(), cli.relay_addr),
+            ],
+            timeout: std::time::Duration::from_secs(5),
+            test_portmap: true,
+            test_ipv6: true,
+            local_port: 0,
+        };
+        let report = wzp_client::netcheck::run_netcheck(&config).await;
+        print!("{}", wzp_client::netcheck::format_report(&report));
+        return Ok(());
+    }
+
     // --version-check: query relay version over QUIC and exit
     if cli.version_check {
         let client_config = wzp_transport::client_config();
@@ -424,6 +446,7 @@ async fn run_silence(transport: Arc<wzp_transport::QuinnTransport>) -> anyhow::R
     info!(total_source, total_repair, total_bytes, "done — closing");
     let hangup = wzp_proto::SignalMessage::Hangup {
         reason: wzp_proto::HangupReason::Normal,
+        call_id: None,
     };
     transport.send_signal(&hangup).await.ok();
     transport.close().await?;
@@ -575,6 +598,7 @@ async fn run_file_mode(
     // Send Hangup signal so the relay knows we're done
     let hangup = wzp_proto::SignalMessage::Hangup {
         reason: wzp_proto::HangupReason::Normal,
+        call_id: None,
     };
     transport.send_signal(&hangup).await.ok();
 
@@ -626,11 +650,21 @@ async fn run_live(transport: Arc<wzp_transport::QuinnTransport>) -> anyhow::Resu
         .spawn(move || {
             let config = CallConfig::default();
             let mut encoder = CallEncoder::new(&config);
+            let mut frame = vec![0i16; FRAME_SAMPLES];
             loop {
-                let frame = match capture.read_frame() {
-                    Some(f) => f,
-                    None => break,
-                };
+                // Pull a full 20 ms frame from the capture ring. The ring
+                // may return a partial read when the CPAL callback hasn't
+                // produced enough samples yet — keep reading until we
+                // accumulate a whole frame, sleeping briefly on empty
+                // returns so we don't hot-spin the CPU.
+                let mut filled = 0usize;
+                while filled < FRAME_SAMPLES {
+                    let n = capture.ring().read(&mut frame[filled..]);
+                    filled += n;
+                    if n == 0 {
+                        std::thread::sleep(std::time::Duration::from_millis(2));
+                    }
+                }
                 let packets = match encoder.encode_frame(&frame) {
                     Ok(p) => p,
                     Err(e) => {
@@ -661,7 +695,13 @@ async fn run_live(transport: Arc<wzp_transport::QuinnTransport>) -> anyhow::Resu
                     // Repair packets feed the FEC decoder but don't produce audio.
                     if !is_repair {
                         if let Some(_n) = decoder.decode_next(&mut pcm_buf) {
-                            playback.write_frame(&pcm_buf);
+                            // Push the decoded frame into the playback
+                            // ring. The CPAL output callback drains from
+                            // here on its own clock; if the ring is full
+                            // (rare in CLI live mode) the write returns
+                            // a short count and the tail is dropped,
+                            // which is the correct real-time behavior.
+                            playback.ring().write(&pcm_buf);
                         }
                     }
                 }
@@ -731,7 +771,7 @@ async fn run_signal_mode(
         Some(SignalMessage::RegisterPresenceAck { success: true, .. }) => {
             info!(fingerprint = %fp, "registered on relay — waiting for calls");
         }
-        Some(SignalMessage::RegisterPresenceAck { success: false, error }) => {
+        Some(SignalMessage::RegisterPresenceAck { success: false, error, .. }) => {
             anyhow::bail!("registration failed: {}", error.unwrap_or_default());
         }
         other => {
@@ -754,13 +794,18 @@ async fn run_signal_mode(
             ephemeral_pub: [0u8; 32], // Phase 1: not used for key exchange
             signature: vec![],
             supported_profiles: vec![wzp_proto::QualityProfile::GOOD],
+            // CLI client doesn't attempt hole-punching; always
+            // relay-path.
+            caller_reflexive_addr: None,
+            caller_local_addrs: Vec::new(),
+            caller_mapped_addr: None,
+            caller_build_version: None,
         }).await?;
     }
 
     // Signal recv loop — handle incoming signals
     let signal_transport = transport.clone();
     let relay = relay_addr;
-    let my_fp = fp.clone();
     let my_seed = seed.0;
 
     loop {
@@ -784,12 +829,18 @@ async fn run_signal_mode(
                         ephemeral_pub: None,
                         signature: None,
                         chosen_profile: Some(wzp_proto::QualityProfile::GOOD),
+                        // CLI auto-accept uses generic (privacy) mode,
+                        // so callee addr stays hidden from the caller.
+                        callee_reflexive_addr: None,
+                        callee_local_addrs: Vec::new(),
+                        callee_mapped_addr: None,
+                        callee_build_version: None,
                     }).await;
                 }
                 SignalMessage::DirectCallAnswer { call_id, accept_mode, .. } => {
                     info!(call_id = %call_id, mode = ?accept_mode, "call answered");
                 }
-                SignalMessage::CallSetup { call_id, room, relay_addr: setup_relay } => {
+                SignalMessage::CallSetup { call_id, room, relay_addr: setup_relay, peer_direct_addr: _, peer_local_addrs: _, peer_mapped_addr: _ } => {
                     info!(call_id = %call_id, room = %room, relay = %setup_relay, "call setup — connecting to media room");
 
                     // Connect to the media room
@@ -840,6 +891,7 @@ async fn run_signal_mode(
                                                 info!("hanging up...");
                                                 let _ = signal_transport.send_signal(&SignalMessage::Hangup {
                                                     reason: wzp_proto::HangupReason::Normal,
+                                                    call_id: None,
                                                 }).await;
                                                 break;
                                             }
@@ -856,7 +908,7 @@ async fn run_signal_mode(
                         Err(e) => error!("media connect failed: {e}"),
                     }
                 }
-                SignalMessage::Hangup { reason } => {
+                SignalMessage::Hangup { reason, .. } => {
                     info!(reason = ?reason, "call ended by remote");
                 }
                 SignalMessage::Pong { .. } => {}

crates/wzp-client/src/dual_path.rs

@@ -0,0 +1,960 @@
+//! Phase 3.5 — dual-path QUIC connect race for P2P hole-punching.
+//!
+//! When both peers advertised reflex addrs in the
+//! DirectCallOffer/Answer flow, the relay cross-wires them into
+//! `CallSetup.peer_direct_addr`. This module races a direct QUIC
+//! handshake against the existing relay dial and returns whichever
+//! completes first — with automatic drop of the loser via
+//! `tokio::select!`.
+//!
+//! Role determination is deterministic and symmetric
+//! (`wzp_client::reflect::determine_role`): whichever peer has the
+//! lexicographically smaller reflex addr becomes the **Acceptor**
+//! (listens on a server-capable endpoint), the other becomes the
+//! **Dialer** (dials the peer's addr). Because the rule is
+//! identical on both sides, the Acceptor's inbound QUIC session
+//! and the Dialer's outbound are the SAME connection — no
+//! negotiation needed, no two-conns-per-call confusion.
+//!
+//! Timeout policy:
+//! - Direct path: 2s from the start of `race`. Cone-NAT hole-punch
+//!   typically completes in < 500ms on a LAN; 2s gives us tolerance
+//!   for a single QUIC Initial retry on unreliable networks.
+//! - Relay path: 10s (existing behavior elsewhere in the codebase).
+//! - Overall: `tokio::select!` returns as soon as either succeeds.
+
+use std::net::SocketAddr;
+use std::sync::Arc;
+use std::time::Duration;
+
+use crate::reflect::Role;
+use wzp_transport::QuinnTransport;
+
+/// Which path won the race. Used by the `connect` command for
+/// logging + (in the future) metrics.
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub enum WinningPath {
+    Direct,
+    Relay,
+}
+
+/// Diagnostic info for a single candidate dial attempt.
+#[derive(Debug, Clone, serde::Serialize)]
+pub struct CandidateDiag {
+    pub index: usize,
+    pub addr: String,
+    pub result: String,  // "ok", "skipped:ipv6", "error:..."
+    pub elapsed_ms: Option<u32>,
+}
+
+/// Phase 6: the race now returns BOTH transports (when available)
+/// so the connect command can negotiate with the peer before
+/// committing. The negotiation decides which transport to use
+/// based on whether BOTH sides report `direct_ok = true`.
+pub struct RaceResult {
+    /// The direct P2P transport, if the direct path completed.
+    /// `None` if the direct dial/accept failed or timed out.
+    pub direct_transport: Option<Arc<QuinnTransport>>,
+    /// The relay transport, if the relay dial completed.
+    /// `None` if the relay dial failed (shouldn't happen in
+    /// practice since relay is always reachable).
+    pub relay_transport: Option<Arc<QuinnTransport>>,
+    /// Which future completed first in the local race.
+    /// Informational — the actual path used is decided by the
+    /// Phase 6 negotiation after both sides exchange reports.
+    pub local_winner: WinningPath,
+    /// Per-candidate diagnostic info for debugging.
+    pub candidate_diags: Vec<CandidateDiag>,
+}
+
+/// Attempt a direct QUIC connection to the peer in parallel with
+/// the relay dial and return the winning `QuinnTransport`.
+///
+/// `role` selects the direction of the direct attempt:
+/// - `Role::Acceptor` creates a server-capable endpoint and waits
+///   for the peer to dial in.
+/// - `Role::Dialer` creates a client-only endpoint and dials
+///   `peer_direct_addr`.
+///
+/// The relay path is always attempted in parallel as a fallback so
+/// the race ALWAYS produces a working transport unless both paths
+/// genuinely fail (network partition). Returns
+/// `Err(anyhow::anyhow!(...))` if both paths fail within the
+/// timeout.
+/// Phase 5.5 candidate bundle — full ICE-ish candidate list for
+/// the peer. The race tries them all in parallel alongside the
+/// relay path. At minimum this should contain the peer's
+/// server-reflexive address; `local_addrs` carries LAN host
+/// candidates gathered from their physical interfaces.
+///
+/// Empty is valid: the D-role has nothing to dial and the race
+/// reduces to "relay only" + (if A-role) accepting on the
+/// shared endpoint.
+#[derive(Debug, Clone, Default)]
+pub struct PeerCandidates {
+    /// Peer's server-reflexive address (Phase 3). `None` if the
+    /// peer didn't advertise one.
+    pub reflexive: Option<SocketAddr>,
+    /// Peer's LAN host addresses (Phase 5.5). Tried first on
+    /// same-LAN pairs — direct dials to these bypass the NAT
+    /// entirely.
+    pub local: Vec<SocketAddr>,
+    /// Phase 8 (Tailscale-inspired): peer's port-mapped external
+    /// address from NAT-PMP/PCP/UPnP. When the router supports
+    /// port mapping, this gives a stable external address even
+    /// behind symmetric NATs.
+    pub mapped: Option<SocketAddr>,
+}
+
+impl PeerCandidates {
+    /// Flatten into the list of addrs the D-role should dial.
+    /// Order: LAN host candidates first (fastest when they
+    /// work), then port-mapped (stable even behind symmetric
+    /// NATs), then reflexive (covers the non-LAN case).
+    pub fn dial_order(&self) -> Vec<SocketAddr> {
+        let mut out = Vec::with_capacity(self.local.len() + 2);
+        out.extend(self.local.iter().copied());
+        // Port-mapped address goes before reflexive — it's
+        // more reliable on symmetric NATs where the reflexive
+        // addr might not match what the peer actually sees.
+        if let Some(a) = self.mapped {
+            if !out.contains(&a) {
+                out.push(a);
+            }
+        }
+        if let Some(a) = self.reflexive {
+            if !out.contains(&a) {
+                out.push(a);
+            }
+        }
+        out
+    }
+
+    /// Smart dial order: filters out candidates that can't possibly
+    /// work given our own reflexive address.
+    ///
+    /// - **LAN candidates**: only included if peer's public IP
+    ///   matches ours (same network). Private IPs are unreachable
+    ///   cross-network.
+    /// - **IPv6 candidates**: stripped entirely (Phase 7 disabled).
+    /// - **Reflexive + mapped**: always included.
+    pub fn smart_dial_order(&self, own_reflexive: Option<&SocketAddr>) -> Vec<SocketAddr> {
+        let own_public_ip = own_reflexive.map(|a| a.ip());
+        let peer_public_ip = self.reflexive.map(|a| a.ip());
+        let same_network = match (own_public_ip, peer_public_ip) {
+            (Some(a), Some(b)) => a == b,
+            _ => false,
+        };
+
+        let mut out = Vec::with_capacity(self.local.len() + 2);
+
+        // LAN candidates only when on the same network.
+        if same_network {
+            for addr in &self.local {
+                if !addr.is_ipv6() {
+                    out.push(*addr);
+                }
+            }
+        }
+
+        // Port-mapped (always useful — it's a public addr).
+        if let Some(a) = self.mapped {
+            if !a.is_ipv6() && !out.contains(&a) {
+                out.push(a);
+            }
+        }
+
+        // Reflexive (always useful — it's the peer's public addr).
+        if let Some(a) = self.reflexive {
+            if !a.is_ipv6() && !out.contains(&a) {
+                out.push(a);
+            }
+        }
+
+        out
+    }
+
+    /// Is there anything for the D-role to dial? If not, the
+    /// race reduces to relay-only.
+    pub fn is_empty(&self) -> bool {
+        self.reflexive.is_none() && self.local.is_empty() && self.mapped.is_none()
+    }
+}
+
+#[allow(clippy::too_many_arguments)]
+pub async fn race(
+    role: Role,
+    peer_candidates: PeerCandidates,
+    relay_addr: SocketAddr,
+    room_sni: String,
+    call_sni: String,
+    // Our own reflexive address — used to filter LAN candidates
+    // that can't work cross-network.
+    own_reflexive: Option<SocketAddr>,
+    // Phase 5: when `Some`, reuse this endpoint for BOTH the
+    // direct-path branch AND the relay dial. Pass the signal
+    // endpoint. The endpoint MUST be server-capable (created
+    // with a server config) for the A-role accept branch to
+    // work.
+    //
+    // When `None`, falls back to fresh endpoints per role.
+    // Used by tests.
+    shared_endpoint: Option<wzp_transport::Endpoint>,
+    // Phase 7: dedicated IPv6 endpoint with IPV6_V6ONLY=1.
+    // When `Some`, A-role accepts on both v4+v6, D-role routes
+    // each candidate to its matching-AF endpoint. When `None`,
+    // IPv6 candidates are skipped (IPv4-only, pre-Phase-7).
+    ipv6_endpoint: Option<wzp_transport::Endpoint>,
+) -> anyhow::Result<RaceResult> {
+    // Rustls provider must be installed before any quinn endpoint
+    // is created. Install attempt is idempotent.
+    let _ = rustls::crypto::ring::default_provider().install_default();
+
+    // Shared diagnostic collector for per-candidate results.
+    let diags_collector: Arc<std::sync::Mutex<Vec<CandidateDiag>>> =
+        Arc::new(std::sync::Mutex::new(Vec::new()));
+
+    // Build the direct-path endpoint + future based on role.
+    //
+    // A-role: one accept future on the shared endpoint. The
+    //   first incoming QUIC connection wins — we don't care
+    //   which peer candidate the dialer used to reach us.
+    //
+    // D-role: N parallel dial futures, one per peer candidate
+    //   (all LAN host addrs + the reflex addr), consolidated
+    //   into a single direct_fut via FuturesUnordered-style
+    //   "first OK wins" semantics. The first successful dial
+    //   becomes the direct path; the losers are dropped (quinn
+    //   will abort the in-flight handshakes via the dropped
+    //   Connecting futures).
+    //
+    // Either way, direct_fut resolves to a single QuinnTransport
+    // (or an error) and is raced against the relay_fut by the
+    // outer tokio::select!.
+    let direct_ep: wzp_transport::Endpoint;
+    let direct_fut: std::pin::Pin<
+        Box<dyn std::future::Future<Output = anyhow::Result<QuinnTransport>> + Send>,
+    >;
+
+    match role {
+        Role::Acceptor => {
+            let ep = match shared_endpoint.clone() {
+                Some(ep) => {
+                    tracing::info!(
+                        local_addr = ?ep.local_addr().ok(),
+                        "dual_path: A-role reusing shared endpoint for accept"
+                    );
+                    ep
+                }
+                None => {
+                    let (sc, _cert_der) = wzp_transport::server_config();
+                    // 0.0.0.0:0 = IPv4 socket. [::]:0 dual-stack was
+                    // tried but breaks on Android devices where
+                    // IPV6_V6ONLY=1 (default on some kernels) —
+                    // IPv4 candidates silently fail. IPv6 host
+                    // candidates are skipped for now; they need a
+                    // dedicated IPv6 socket alongside the v4 one
+                    // (like WebRTC's dual-socket approach).
+                    let bind: SocketAddr = "0.0.0.0:0".parse().unwrap();
+                    let fresh = wzp_transport::create_endpoint(bind, Some(sc))?;
+                    tracing::info!(
+                        local_addr = ?fresh.local_addr().ok(),
+                        "dual_path: A-role fresh endpoint up, awaiting peer dial"
+                    );
+                    fresh
+                }
+            };
+            let ep_for_fut = ep.clone();
+            // Phase 7: IPv6 accept temporarily disabled (same reason
+            // as dial — IPv6 connections die on datagram send).
+            // Accept on IPv4 shared endpoint only.
+            let _v6_ep_unused = ipv6_endpoint.clone();
+            // Collect peer addrs for NAT tickle (Acceptor-side).
+            let tickle_addrs: Vec<SocketAddr> = peer_candidates
+                .smart_dial_order(own_reflexive.as_ref())
+                .into_iter()
+                .filter(|a| !a.ip().is_loopback() && !a.ip().is_unspecified())
+                .collect();
+            direct_fut = Box::pin(async move {
+                // NAT tickle: send a small UDP packet to each of the
+                // Dialer's candidate addresses FROM our shared endpoint.
+                // This opens our NAT's pinhole for return traffic from
+                // those IPs — critical for address-restricted NATs that
+                // only allow inbound from IPs they've seen outbound
+                // traffic to. Without this, the Dialer's QUIC Initial
+                // gets dropped by our NAT.
+                if !tickle_addrs.is_empty() {
+                    if let Ok(local_addr) = ep_for_fut.local_addr() {
+                        // Send a tickle to each peer candidate address
+                        // to open our NAT for return traffic from that IP.
+                        //
+                        // We use a socket2 socket with SO_REUSEADDR +
+                        // SO_REUSEPORT on the SAME port as the quinn
+                        // endpoint. This is necessary because quinn
+                        // already holds the port — a plain bind() would
+                        // fail with EADDRINUSE.
+                        let tickle_result: Result<(), String> = (|| {
+                            use std::net::UdpSocket as StdUdpSocket;
+                            let sock = socket2::Socket::new(
+                                socket2::Domain::IPV4,
+                                socket2::Type::DGRAM,
+                                Some(socket2::Protocol::UDP),
+                            ).map_err(|e| format!("socket: {e}"))?;
+                            sock.set_reuse_address(true).map_err(|e| format!("reuseaddr: {e}"))?;
+                            // macOS/BSD/Linux also need SO_REUSEPORT
+                            #[cfg(any(target_os = "macos", target_os = "linux", target_os = "android"))]
+                            {
+                                // socket2 exposes set_reuse_port on unix
+                                unsafe {
+                                    let optval: libc::c_int = 1;
+                                    libc::setsockopt(
+                                        std::os::unix::io::AsRawFd::as_raw_fd(&sock),
+                                        libc::SOL_SOCKET,
+                                        libc::SO_REUSEPORT,
+                                        &optval as *const _ as *const libc::c_void,
+                                        std::mem::size_of::<libc::c_int>() as libc::socklen_t,
+                                    );
+                                }
+                            }
+                            sock.set_nonblocking(true).map_err(|e| format!("nonblock: {e}"))?;
+                            let bind_addr: SocketAddr = SocketAddr::new(
+                                std::net::IpAddr::V4(std::net::Ipv4Addr::UNSPECIFIED),
+                                local_addr.port(),
+                            );
+                            sock.bind(&bind_addr.into()).map_err(|e| format!("bind :{}: {e}", local_addr.port()))?;
+                            let std_sock: StdUdpSocket = sock.into();
+                            for addr in &tickle_addrs {
+                                let _ = std_sock.send_to(&[0u8; 1], addr);
+                                tracing::info!(
+                                    %addr,
+                                    local_port = local_addr.port(),
+                                    "dual_path: A-role sent NAT tickle"
+                                );
+                            }
+                            Ok(())
+                        })();
+                        if let Err(e) = tickle_result {
+                            tracing::warn!(error = %e, "dual_path: A-role NAT tickle failed");
+                        }
+                    }
+                }
+
+                // Accept loop: retry if we get a stale/closed
+                // connection from a previous call. Max 3 retries
+                // to avoid spinning until the race timeout.
+                const MAX_STALE: usize = 3;
+                let mut stale_count: usize = 0;
+                loop {
+                    let conn = wzp_transport::accept(&ep_for_fut)
+                        .await
+                        .map_err(|e| anyhow::anyhow!("direct accept: {e}"))?;
+
+                    if let Some(reason) = conn.close_reason() {
+                        // Explicitly close so the peer gets a
+                        // close frame instead of idle timeout.
+                        conn.close(0u32.into(), b"stale");
+                        stale_count += 1;
+                        tracing::warn!(
+                            remote = %conn.remote_address(),
+                            stable_id = conn.stable_id(),
+                            stale_count,
+                            ?reason,
+                            "dual_path: A-role skipping stale connection"
+                        );
+                        if stale_count >= MAX_STALE {
+                            return Err(anyhow::anyhow!(
+                                "A-role: {stale_count} stale connections, aborting"
+                            ));
+                        }
+                        continue;
+                    }
+
+                    let has_dgram = conn.max_datagram_size().is_some();
+                    tracing::info!(
+                        remote = %conn.remote_address(),
+                        stable_id = conn.stable_id(),
+                        has_dgram,
+                        "dual_path: A-role accepted direct connection"
+                    );
+
+                    break Ok(QuinnTransport::new(conn));
+                }
+            });
+            direct_ep = ep;
+        }
+        Role::Dialer => {
+            let ep = match shared_endpoint.clone() {
+                Some(ep) => {
+                    tracing::info!(
+                        local_addr = ?ep.local_addr().ok(),
+                        candidates = ?peer_candidates.dial_order(),
+                        "dual_path: D-role reusing shared endpoint to dial peer candidates"
+                    );
+                    ep
+                }
+                None => {
+                    // 0.0.0.0:0 = IPv4 socket. [::]:0 dual-stack was
+                    // tried but breaks on Android devices where
+                    // IPV6_V6ONLY=1 (default on some kernels) —
+                    // IPv4 candidates silently fail. IPv6 host
+                    // candidates are skipped for now; they need a
+                    // dedicated IPv6 socket alongside the v4 one
+                    // (like WebRTC's dual-socket approach).
+                    let bind: SocketAddr = "0.0.0.0:0".parse().unwrap();
+                    let fresh = wzp_transport::create_endpoint(bind, None)?;
+                    tracing::info!(
+                        local_addr = ?fresh.local_addr().ok(),
+                        candidates = ?peer_candidates.dial_order(),
+                        "dual_path: D-role fresh endpoint up, dialing peer candidates"
+                    );
+                    fresh
+                }
+            };
+            let ep_for_fut = ep.clone();
+            let _v6_ep_for_dial = ipv6_endpoint.clone();
+            let dial_order = peer_candidates.smart_dial_order(own_reflexive.as_ref());
+            let sni = call_sni.clone();
+            let diags = diags_collector.clone();
+            direct_fut = Box::pin(async move {
+                if dial_order.is_empty() {
+                    // No candidates — the race reduces to
+                    // relay-only. Surface a stable error so the
+                    // outer select falls through to relay_fut
+                    // without a spurious "direct failed" warning.
+                    // Use a pending future that never resolves so
+                    // the select's "other side wins" branch is
+                    // the natural outcome.
+                    std::future::pending::<anyhow::Result<QuinnTransport>>().await
+                } else {
+                    // Fan out N parallel dials via JoinSet. First
+                    // `Ok` wins; `Err` from a single candidate is
+                    // not fatal — we wait for the others. Only
+                    // when ALL have failed do we return Err.
+                    let mut set = tokio::task::JoinSet::new();
+                    for (idx, candidate) in dial_order.iter().enumerate() {
+                        // Phase 7: route each candidate to the
+                        // endpoint matching its address family.
+                        let candidate = *candidate;
+                        // Phase 7: IPv6 dials temporarily disabled.
+                        // IPv6 QUIC handshakes succeed but the
+                        // connection dies immediately on datagram
+                        // send ("connection lost"). Root cause is
+                        // likely router-level IPv6 UDP filtering.
+                        // Re-enable once IPv6 datagram delivery is
+                        // verified on target networks.
+                        if candidate.is_ipv6() {
+                            tracing::info!(
+                                %candidate,
+                                candidate_idx = idx,
+                                "dual_path: skipping IPv6 candidate (disabled)"
+                            );
+                            if let Ok(mut d) = diags.lock() {
+                                d.push(CandidateDiag {
+                                    index: idx,
+                                    addr: candidate.to_string(),
+                                    result: "skipped:ipv6".into(),
+                                    elapsed_ms: None,
+                                });
+                            }
+                            continue;
+                        }
+                        let ep = ep_for_fut.clone();
+                        let client_cfg = wzp_transport::client_config();
+                        let sni = sni.clone();
+                        let diags_inner = diags.clone();
+                        set.spawn(async move {
+                            let start = std::time::Instant::now();
+                            tracing::info!(
+                                %candidate,
+                                candidate_idx = idx,
+                                "dual_path: dialing candidate"
+                            );
+                            let result = wzp_transport::connect(
+                                &ep,
+                                candidate,
+                                &sni,
+                                client_cfg,
+                            )
+                            .await;
+                            let elapsed = start.elapsed().as_millis() as u32;
+                            let diag_result = match &result {
+                                Ok(_) => "ok".to_string(),
+                                Err(e) => format!("error:{e}"),
+                            };
+                            if let Ok(mut d) = diags_inner.lock() {
+                                d.push(CandidateDiag {
+                                    index: idx,
+                                    addr: candidate.to_string(),
+                                    result: diag_result,
+                                    elapsed_ms: Some(elapsed),
+                                });
+                            }
+                            (idx, candidate, result)
+                        });
+                    }
+                    let mut last_err: Option<String> = None;
+                    while let Some(join_res) = set.join_next().await {
+                        let (idx, candidate, dial_res) = match join_res {
+                            Ok(t) => t,
+                            Err(e) => {
+                                last_err = Some(format!("join {e}"));
+                                continue;
+                            }
+                        };
+                        match dial_res {
+                            Ok(conn) => {
+                                tracing::info!(
+                                    %candidate,
+                                    candidate_idx = idx,
+                                    remote = %conn.remote_address(),
+                                    stable_id = conn.stable_id(),
+                                    "dual_path: direct dial succeeded on candidate"
+                                );
+                                // Abort the remaining in-flight
+                                // dials so they don't complete
+                                // and leak QUIC sessions.
+                                set.abort_all();
+                                return Ok(QuinnTransport::new(conn));
+                            }
+                            Err(e) => {
+                                tracing::info!(
+                                    %candidate,
+                                    candidate_idx = idx,
+                                    error = %e,
+                                    "dual_path: direct dial failed, trying others"
+                                );
+                                last_err = Some(format!("candidate {candidate}: {e}"));
+                            }
+                        }
+                    }
+                    Err(anyhow::anyhow!(
+                        "all {} direct candidates failed; last: {}",
+                        dial_order.len(),
+                        last_err.unwrap_or_else(|| "n/a".into())
+                    ))
+                }
+            });
+            direct_ep = ep;
+        }
+    }
+
+    // Relay path: classic dial to the relay's media room. Phase 5:
+    // reuse the shared endpoint here too so MikroTik-style NATs
+    // keep a stable external port across all flows from this
+    // client. Falls back to a fresh endpoint when not shared.
+    let relay_ep = match shared_endpoint.clone() {
+        Some(ep) => ep,
+        None => {
+            let relay_bind: SocketAddr = "[::]:0".parse().unwrap();
+            wzp_transport::create_endpoint(relay_bind, None)?
+        }
+    };
+    let relay_ep_for_fut = relay_ep.clone();
+    let relay_client_cfg = wzp_transport::client_config();
+    let relay_sni = room_sni.clone();
+    // Phase 5.5 direct-path head-start: hold the relay dial for
+    // 500ms before attempting it. On same-LAN cone-NAT pairs the
+    // direct dial finishes in ~30-100ms, so giving direct a 500ms
+    // head start means direct reliably wins when it's going to
+    // work at all. The worst case adds 500ms to the fall-back-
+    // to-relay scenario, which is imperceptible for users on
+    // setups where direct isn't available anyway.
+    //
+    // Prior behavior (immediate race) caused the relay to win
+    // ~105ms races on a MikroTik LAN because:
+    //   - Acceptor role's direct_fut = accept() can only fire
+    //     when the peer has completed its outbound LAN dial
+    //   - Dialer role's parallel LAN dials need the peer's
+    //     CallSetup processed + the race started on the other
+    //     side before they can reach us
+    //   - Meanwhile relay_fut is a plain dial that completes in
+    //     whatever the client→relay RTT is (often <100ms)
+    //
+    // The 500ms head start is the minimum that empirically makes
+    // same-LAN direct reliably beat relay, without penalizing
+    // users who genuinely need the relay path.
+    const DIRECT_HEAD_START: Duration = Duration::from_millis(500);
+    let relay_fut = async move {
+        tokio::time::sleep(DIRECT_HEAD_START).await;
+        let conn =
+            wzp_transport::connect(&relay_ep_for_fut, relay_addr, &relay_sni, relay_client_cfg)
+                .await
+                .map_err(|e| anyhow::anyhow!("relay dial: {e}"))?;
+        Ok::<_, anyhow::Error>(QuinnTransport::new(conn))
+    };
+
+    // Phase 6: run both paths concurrently via tokio::spawn and
+    // collect BOTH results. The old tokio::select! approach dropped
+    // the loser, which meant the connect command couldn't negotiate
+    // with the peer — it had to commit to whichever path won locally.
+    //
+    // Now we spawn both as tasks, wait for the first to complete
+    // (that determines `local_winner`), then give the loser a short
+    // grace period to also complete. The connect command gets a
+    // RaceResult with both transports (when available) and uses the
+    // Phase 6 MediaPathReport exchange to decide which one to
+    // actually use for media.
+    let smart_order = peer_candidates.smart_dial_order(own_reflexive.as_ref());
+    tracing::info!(
+        ?role,
+        raw_candidates = ?peer_candidates.dial_order(),
+        filtered_candidates = ?smart_order,
+        ?own_reflexive,
+        %relay_addr,
+        "dual_path: racing direct vs relay"
+    );
+
+    let mut direct_task = tokio::spawn(
+        tokio::time::timeout(Duration::from_secs(4), direct_fut),
+    );
+    let mut relay_task = tokio::spawn(async move {
+        // Keep the 500ms head start so direct has a chance
+        tokio::time::sleep(Duration::from_millis(500)).await;
+        tokio::time::timeout(Duration::from_secs(5), relay_fut).await
+    });
+
+    // Wait for the first one to complete. This tells us the
+    // local_winner — but we DON'T commit to it yet. Phase 6
+    // negotiation decides the actual path.
+    let (mut direct_result, mut relay_result): (
+        Option<anyhow::Result<QuinnTransport>>,
+        Option<anyhow::Result<QuinnTransport>>,
+    ) = (None, None);
+
+    let local_winner;
+
+    tokio::select! {
+        biased;
+        d = &mut direct_task => {
+            match d {
+                Ok(Ok(Ok(t))) => {
+                    tracing::info!("dual_path: direct completed first");
+                    direct_result = Some(Ok(t));
+                    local_winner = WinningPath::Direct;
+                }
+                Ok(Ok(Err(e))) => {
+                    tracing::warn!(error = %e, "dual_path: direct failed");
+                    direct_result = Some(Err(anyhow::anyhow!("{e}")));
+                    local_winner = WinningPath::Relay; // direct failed → relay is our only hope
+                }
+                Ok(Err(_)) => {
+                    tracing::warn!("dual_path: direct timed out (4s)");
+                    direct_result = Some(Err(anyhow::anyhow!("direct timeout")));
+                    local_winner = WinningPath::Relay;
+                    // Record timeout diag for candidates that were
+                    // still in-flight when the timeout fired.
+                    if let Ok(mut d) = diags_collector.lock() {
+                        let recorded_indices: std::collections::HashSet<usize> =
+                            d.iter().map(|diag| diag.index).collect();
+                        for (idx, addr) in smart_order.iter().enumerate() {
+                            if !recorded_indices.contains(&idx) {
+                                d.push(CandidateDiag {
+                                    index: idx,
+                                    addr: addr.to_string(),
+                                    result: "timeout:4s".into(),
+                                    elapsed_ms: Some(4000),
+                                });
+                            }
+                        }
+                    }
+                }
+                Err(e) => {
+                    tracing::warn!(error = %e, "dual_path: direct task panicked");
+                    direct_result = Some(Err(anyhow::anyhow!("direct task panic")));
+                    local_winner = WinningPath::Relay;
+                }
+            }
+        }
+        r = &mut relay_task => {
+            match r {
+                Ok(Ok(Ok(t))) => {
+                    tracing::info!("dual_path: relay completed first");
+                    relay_result = Some(Ok(t));
+                    local_winner = WinningPath::Relay;
+                }
+                Ok(Ok(Err(e))) => {
+                    tracing::warn!(error = %e, "dual_path: relay failed");
+                    relay_result = Some(Err(anyhow::anyhow!("{e}")));
+                    local_winner = WinningPath::Direct;
+                }
+                Ok(Err(_)) => {
+                    tracing::warn!("dual_path: relay timed out");
+                    relay_result = Some(Err(anyhow::anyhow!("relay timeout")));
+                    local_winner = WinningPath::Direct;
+                }
+                Err(e) => {
+                    relay_result = Some(Err(anyhow::anyhow!("relay task panic: {e}")));
+                    local_winner = WinningPath::Direct;
+                }
+            }
+        }
+    }
+
+    // Give the loser a short grace period (1s) to also complete.
+    // If it does, we have both transports for Phase 6 negotiation.
+    // If it doesn't, we still proceed with just the winner.
+    if direct_result.is_none() {
+        match tokio::time::timeout(Duration::from_secs(1), direct_task).await {
+            Ok(Ok(Ok(Ok(t)))) => { direct_result = Some(Ok(t)); }
+            Ok(Ok(Ok(Err(e)))) => { direct_result = Some(Err(anyhow::anyhow!("{e}"))); }
+            _ => {
+                direct_result = Some(Err(anyhow::anyhow!("direct: no result in grace period")));
+                // Fill timeout diags for candidates that never reported.
+                if let Ok(mut d) = diags_collector.lock() {
+                    let recorded: std::collections::HashSet<usize> =
+                        d.iter().map(|diag| diag.index).collect();
+                    for (idx, addr) in smart_order.iter().enumerate() {
+                        if !recorded.contains(&idx) {
+                            d.push(CandidateDiag {
+                                index: idx,
+                                addr: addr.to_string(),
+                                result: "timeout:grace".into(),
+                                elapsed_ms: None,
+                            });
+                        }
+                    }
+                }
+            }
+        }
+    }
+    if relay_result.is_none() {
+        match tokio::time::timeout(Duration::from_secs(1), relay_task).await {
+            Ok(Ok(Ok(Ok(t)))) => { relay_result = Some(Ok(t)); }
+            Ok(Ok(Ok(Err(e)))) => { relay_result = Some(Err(anyhow::anyhow!("{e}"))); }
+            _ => { relay_result = Some(Err(anyhow::anyhow!("relay: no result in grace period"))); }
+        }
+    }
+
+    let direct_ok = direct_result.as_ref().map(|r| r.is_ok()).unwrap_or(false);
+    let relay_ok = relay_result.as_ref().map(|r| r.is_ok()).unwrap_or(false);
+
+    tracing::info!(
+        ?local_winner,
+        direct_ok,
+        relay_ok,
+        "dual_path: race finished, both results collected for Phase 6 negotiation"
+    );
+
+    if !direct_ok && !relay_ok {
+        return Err(anyhow::anyhow!("both paths failed: no media transport available"));
+    }
+
+    let _ = (direct_ep, relay_ep, ipv6_endpoint);
+
+    let candidate_diags = diags_collector.lock()
+        .map(|d| d.clone())
+        .unwrap_or_default();
+
+    Ok(RaceResult {
+        direct_transport: direct_result
+            .and_then(|r| r.ok())
+            .map(|t| Arc::new(t)),
+        relay_transport: relay_result
+            .and_then(|r| r.ok())
+            .map(|t| Arc::new(t)),
+        local_winner,
+        candidate_diags,
+    })
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn peer_candidates_dial_order_all_types() {
+        let candidates = PeerCandidates {
+            reflexive: Some("203.0.113.5:4433".parse().unwrap()),
+            local: vec![
+                "192.168.1.10:4433".parse().unwrap(),
+                "10.0.0.5:4433".parse().unwrap(),
+            ],
+            mapped: Some("198.51.100.42:12345".parse().unwrap()),
+        };
+
+        let order = candidates.dial_order();
+        // Order: local first, then mapped, then reflexive
+        assert_eq!(order.len(), 4);
+        assert_eq!(order[0], "192.168.1.10:4433".parse::<SocketAddr>().unwrap());
+        assert_eq!(order[1], "10.0.0.5:4433".parse::<SocketAddr>().unwrap());
+        assert_eq!(order[2], "198.51.100.42:12345".parse::<SocketAddr>().unwrap());
+        assert_eq!(order[3], "203.0.113.5:4433".parse::<SocketAddr>().unwrap());
+    }
+
+    #[test]
+    fn peer_candidates_dial_order_no_mapped() {
+        let candidates = PeerCandidates {
+            reflexive: Some("203.0.113.5:4433".parse().unwrap()),
+            local: vec!["192.168.1.10:4433".parse().unwrap()],
+            mapped: None,
+        };
+
+        let order = candidates.dial_order();
+        assert_eq!(order.len(), 2);
+        assert_eq!(order[0], "192.168.1.10:4433".parse::<SocketAddr>().unwrap());
+        assert_eq!(order[1], "203.0.113.5:4433".parse::<SocketAddr>().unwrap());
+    }
+
+    #[test]
+    fn peer_candidates_dial_order_only_mapped() {
+        let candidates = PeerCandidates {
+            reflexive: None,
+            local: vec![],
+            mapped: Some("198.51.100.42:12345".parse().unwrap()),
+        };
+
+        let order = candidates.dial_order();
+        assert_eq!(order.len(), 1);
+        assert_eq!(order[0], "198.51.100.42:12345".parse::<SocketAddr>().unwrap());
+    }
+
+    #[test]
+    fn peer_candidates_dial_order_dedup_mapped_equals_reflexive() {
+        let addr: SocketAddr = "203.0.113.5:4433".parse().unwrap();
+        let candidates = PeerCandidates {
+            reflexive: Some(addr),
+            local: vec![],
+            mapped: Some(addr), // same as reflexive
+        };
+
+        let order = candidates.dial_order();
+        // Should be deduped to 1
+        assert_eq!(order.len(), 1);
+        assert_eq!(order[0], addr);
+    }
+
+    #[test]
+    fn peer_candidates_dial_order_dedup_mapped_in_local() {
+        let addr: SocketAddr = "192.168.1.10:4433".parse().unwrap();
+        let candidates = PeerCandidates {
+            reflexive: None,
+            local: vec![addr],
+            mapped: Some(addr), // same as a local addr
+        };
+
+        let order = candidates.dial_order();
+        assert_eq!(order.len(), 1);
+        assert_eq!(order[0], addr);
+    }
+
+    #[test]
+    fn peer_candidates_is_empty() {
+        let empty = PeerCandidates::default();
+        assert!(empty.is_empty());
+
+        let with_reflexive = PeerCandidates {
+            reflexive: Some("1.2.3.4:5".parse().unwrap()),
+            ..Default::default()
+        };
+        assert!(!with_reflexive.is_empty());
+
+        let with_local = PeerCandidates {
+            local: vec!["10.0.0.1:5".parse().unwrap()],
+            ..Default::default()
+        };
+        assert!(!with_local.is_empty());
+
+        let with_mapped = PeerCandidates {
+            mapped: Some("1.2.3.4:5".parse().unwrap()),
+            ..Default::default()
+        };
+        assert!(!with_mapped.is_empty());
+    }
+
+    #[test]
+    fn peer_candidates_empty_dial_order() {
+        let empty = PeerCandidates::default();
+        assert!(empty.dial_order().is_empty());
+    }
+
+    #[test]
+    fn winning_path_debug() {
+        // Just verify Debug impl doesn't panic
+        let _ = format!("{:?}", WinningPath::Direct);
+        let _ = format!("{:?}", WinningPath::Relay);
+    }
+
+    // ── smart_dial_order tests ─────────────────────────────────
+
+    #[test]
+    fn smart_dial_order_same_network_includes_lan() {
+        let candidates = PeerCandidates {
+            reflexive: Some("203.0.113.5:4433".parse().unwrap()),
+            local: vec![
+                "192.168.1.10:4433".parse().unwrap(),
+                "10.0.0.5:4433".parse().unwrap(),
+            ],
+            mapped: None,
+        };
+        let own: SocketAddr = "203.0.113.5:12345".parse().unwrap();
+        let order = candidates.smart_dial_order(Some(&own));
+        // Same public IP → LAN candidates included
+        assert!(order.contains(&"192.168.1.10:4433".parse().unwrap()));
+        assert!(order.contains(&"10.0.0.5:4433".parse().unwrap()));
+        assert!(order.contains(&"203.0.113.5:4433".parse().unwrap()));
+    }
+
+    #[test]
+    fn smart_dial_order_different_network_strips_lan() {
+        let candidates = PeerCandidates {
+            reflexive: Some("150.228.49.65:4433".parse().unwrap()),
+            local: vec![
+                "172.16.81.126:4433".parse().unwrap(),
+                "10.0.0.5:4433".parse().unwrap(),
+            ],
+            mapped: None,
+        };
+        // Different public IP → LAN candidates stripped
+        let own: SocketAddr = "185.115.4.212:12345".parse().unwrap();
+        let order = candidates.smart_dial_order(Some(&own));
+        assert!(!order.contains(&"172.16.81.126:4433".parse().unwrap()));
+        assert!(!order.contains(&"10.0.0.5:4433".parse().unwrap()));
+        // Reflexive still included
+        assert!(order.contains(&"150.228.49.65:4433".parse().unwrap()));
+    }
+
+    #[test]
+    fn smart_dial_order_strips_ipv6() {
+        let candidates = PeerCandidates {
+            reflexive: Some("150.228.49.65:4433".parse().unwrap()),
+            local: vec![
+                "[2a0d:3344:692c::1]:4433".parse().unwrap(),
+                "172.16.81.126:4433".parse().unwrap(),
+            ],
+            mapped: None,
+        };
+        // Same network, but IPv6 should be stripped
+        let own: SocketAddr = "150.228.49.65:5555".parse().unwrap();
+        let order = candidates.smart_dial_order(Some(&own));
+        assert!(!order.iter().any(|a| a.is_ipv6()));
+        assert!(order.contains(&"172.16.81.126:4433".parse().unwrap()));
+    }
+
+    #[test]
+    fn smart_dial_order_no_own_reflexive_strips_lan() {
+        let candidates = PeerCandidates {
+            reflexive: Some("150.228.49.65:4433".parse().unwrap()),
+            local: vec!["172.16.81.126:4433".parse().unwrap()],
+            mapped: Some("198.51.100.42:12345".parse().unwrap()),
+        };
+        // No own reflexive → can't determine same network → strip LAN
+        let order = candidates.smart_dial_order(None);
+        assert!(!order.contains(&"172.16.81.126:4433".parse().unwrap()));
+        assert!(order.contains(&"198.51.100.42:12345".parse().unwrap()));
+        assert!(order.contains(&"150.228.49.65:4433".parse().unwrap()));
+    }
+
+    #[test]
+    fn smart_dial_order_mapped_always_included() {
+        let candidates = PeerCandidates {
+            reflexive: Some("150.228.49.65:4433".parse().unwrap()),
+            local: vec![],
+            mapped: Some("198.51.100.42:12345".parse().unwrap()),
+        };
+        let own: SocketAddr = "185.115.4.212:12345".parse().unwrap();
+        let order = candidates.smart_dial_order(Some(&own));
+        assert_eq!(order.len(), 2); // mapped + reflexive
+        assert!(order.contains(&"198.51.100.42:12345".parse().unwrap()));
+        assert!(order.contains(&"150.228.49.65:4433".parse().unwrap()));
+    }
+}

crates/wzp-client/src/featherchat.rs

@@ -96,6 +96,7 @@ pub fn signal_to_call_type(signal: &SignalMessage) -> CallSignalType {
         SignalMessage::Hangup { .. } => CallSignalType::Hangup,
         SignalMessage::Rekey { .. } => CallSignalType::Offer, // reuse
         SignalMessage::QualityUpdate { .. } => CallSignalType::Offer, // reuse
+        SignalMessage::LossRecoveryUpdate { .. } => CallSignalType::Offer, // reuse (telemetry)
         SignalMessage::Ping { .. } | SignalMessage::Pong { .. } => CallSignalType::Offer,
         SignalMessage::AuthToken { .. } => CallSignalType::Offer,
         SignalMessage::Hold => CallSignalType::Hold,
@@ -119,6 +120,26 @@ pub fn signal_to_call_type(signal: &SignalMessage) -> CallSignalType {
         SignalMessage::CallRinging { .. } => CallSignalType::Ringing,
         SignalMessage::RegisterPresence { .. }
         | SignalMessage::RegisterPresenceAck { .. } => CallSignalType::Offer, // relay-only
+        // NAT reflection is a client↔relay control exchange that
+        // never crosses the featherChat bridge — if it ever reaches
+        // this mapper something is wrong, but we still have to give
+        // an answer. "Offer" is the generic catch-all.
+        SignalMessage::Reflect
+        | SignalMessage::ReflectResponse { .. } => CallSignalType::Offer, // control-plane
+        // Phase 4 cross-relay forwarding envelope — strictly a
+        // relay-to-relay message, never rides the featherChat
+        // bridge. Catch-all mapping for completeness.
+        SignalMessage::FederatedSignalForward { .. } => CallSignalType::Offer,
+        SignalMessage::MediaPathReport { .. } => CallSignalType::Offer, // control-plane
+        SignalMessage::CandidateUpdate { .. } => CallSignalType::IceCandidate, // mid-call re-gather
+        SignalMessage::HardNatProbe { .. } => CallSignalType::IceCandidate, // hard NAT coordination
+        SignalMessage::HardNatBirthdayStart { .. } => CallSignalType::IceCandidate, // birthday attack
+        SignalMessage::UpgradeProposal { .. }
+        | SignalMessage::UpgradeResponse { .. }
+        | SignalMessage::UpgradeConfirm { .. }
+        | SignalMessage::QualityCapability { .. } => CallSignalType::Offer, // quality negotiation
+        SignalMessage::PresenceList { .. } => CallSignalType::Offer, // lobby presence
+        SignalMessage::QualityDirective { .. } => CallSignalType::Offer, // relay-initiated
     }
 }
 
@@ -158,6 +179,7 @@ mod tests {
 
         let hangup = SignalMessage::Hangup {
             reason: wzp_proto::HangupReason::Normal,
+            call_id: None,
         };
         assert!(matches!(signal_to_call_type(&hangup), CallSignalType::Hangup));
 

crates/wzp-client/src/ice_agent.rs

@@ -0,0 +1,444 @@
+//! Phase 8 (Tailscale-inspired): ICE agent for candidate lifecycle
+//! management and mid-call re-gathering.
+//!
+//! The `IceAgent` owns the state of all candidate discovery
+//! mechanisms (STUN, port mapping, host candidates) and provides:
+//!
+//! - `gather()`: initial candidate gathering during call setup
+//! - `re_gather()`: triggered on network change, produces a
+//!   `CandidateUpdate` to send to the peer
+//! - `apply_peer_update()`: processes peer's candidate updates
+//!
+//! This is NOT a full ICE agent (RFC 8445). It's the Tailscale-style
+//! "gather all candidates, race them all in parallel, pick the
+//! winner" approach, adapted for QUIC transport.
+
+use std::net::SocketAddr;
+use std::sync::atomic::{AtomicU32, Ordering};
+use std::time::Duration;
+
+use wzp_proto::SignalMessage;
+
+use crate::dual_path::PeerCandidates;
+use crate::portmap;
+use crate::reflect;
+use crate::stun;
+
+/// All candidates gathered for the local side.
+#[derive(Debug, Clone)]
+pub struct CandidateSet {
+    /// STUN-discovered server-reflexive address.
+    pub reflexive: Option<SocketAddr>,
+    /// LAN host candidates from local interfaces.
+    pub local: Vec<SocketAddr>,
+    /// Port-mapped address from NAT-PMP/PCP/UPnP.
+    pub mapped: Option<SocketAddr>,
+    /// Generation counter (monotonically increasing per call).
+    pub generation: u32,
+}
+
+/// Configuration for the ICE agent.
+#[derive(Debug, Clone)]
+pub struct IceAgentConfig {
+    /// STUN servers to use for reflexive discovery.
+    pub stun_config: stun::StunConfig,
+    /// Whether to attempt port mapping.
+    pub enable_portmap: bool,
+    /// Timeout for each discovery mechanism.
+    pub gather_timeout: Duration,
+    /// The QUIC endpoint's local port (for host candidate pairing).
+    pub local_v4_port: u16,
+    /// Optional IPv6 port.
+    pub local_v6_port: Option<u16>,
+}
+
+impl Default for IceAgentConfig {
+    fn default() -> Self {
+        Self {
+            stun_config: stun::StunConfig::default(),
+            enable_portmap: true,
+            gather_timeout: Duration::from_secs(3),
+            local_v4_port: 0,
+            local_v6_port: None,
+        }
+    }
+}
+
+/// ICE agent managing candidate lifecycle.
+pub struct IceAgent {
+    config: IceAgentConfig,
+    generation: AtomicU32,
+    call_id: String,
+    /// Last-seen peer generation (to filter stale updates).
+    peer_generation: AtomicU32,
+}
+
+impl IceAgent {
+    pub fn new(call_id: String, config: IceAgentConfig) -> Self {
+        Self {
+            config,
+            generation: AtomicU32::new(0),
+            call_id,
+            peer_generation: AtomicU32::new(0),
+        }
+    }
+
+    /// Initial candidate gathering. Runs all discovery mechanisms
+    /// in parallel and returns the full candidate set.
+    pub async fn gather(&self) -> CandidateSet {
+        let generation = self.generation.fetch_add(1, Ordering::Relaxed);
+
+        // Run STUN + port mapping + host candidates in parallel.
+        let stun_fut = stun::discover_reflexive(&self.config.stun_config);
+        let portmap_fut = async {
+            if self.config.enable_portmap && self.config.local_v4_port > 0 {
+                portmap::acquire_port_mapping(self.config.local_v4_port, None)
+                    .await
+                    .ok()
+            } else {
+                None
+            }
+        };
+
+        let (stun_result, portmap_result) = tokio::join!(
+            tokio::time::timeout(self.config.gather_timeout, stun_fut),
+            tokio::time::timeout(self.config.gather_timeout, portmap_fut),
+        );
+
+        let reflexive = stun_result.ok().and_then(|r| r.ok());
+        let mapped = portmap_result
+            .ok()
+            .flatten()
+            .map(|m| m.external_addr);
+        let local = reflect::local_host_candidates(
+            self.config.local_v4_port,
+            self.config.local_v6_port,
+        );
+
+        tracing::info!(
+            generation,
+            reflexive = ?reflexive,
+            mapped = ?mapped,
+            local_count = local.len(),
+            "ice_agent: gathered candidates"
+        );
+
+        CandidateSet {
+            reflexive,
+            local,
+            mapped,
+            generation,
+        }
+    }
+
+    /// Re-gather candidates after a network change. Increments the
+    /// generation counter and returns a `CandidateUpdate` signal
+    /// message to send to the peer.
+    pub async fn re_gather(&self) -> (CandidateSet, SignalMessage) {
+        let candidates = self.gather().await;
+
+        let update = SignalMessage::CandidateUpdate {
+            call_id: self.call_id.clone(),
+            reflexive_addr: candidates.reflexive.map(|a| a.to_string()),
+            local_addrs: candidates.local.iter().map(|a| a.to_string()).collect(),
+            mapped_addr: candidates.mapped.map(|a| a.to_string()),
+            generation: candidates.generation,
+        };
+
+        (candidates, update)
+    }
+
+    /// Process a peer's candidate update. Returns `Some(PeerCandidates)`
+    /// if the update is newer than the last-seen generation, `None`
+    /// if it's stale.
+    pub fn apply_peer_update(
+        &self,
+        update: &SignalMessage,
+    ) -> Option<PeerCandidates> {
+        let (reflexive_addr, local_addrs, mapped_addr, generation) = match update {
+            SignalMessage::CandidateUpdate {
+                reflexive_addr,
+                local_addrs,
+                mapped_addr,
+                generation,
+                ..
+            } => (reflexive_addr, local_addrs, mapped_addr, *generation),
+            _ => return None,
+        };
+
+        // Only accept if newer than last-seen generation.
+        let prev = self.peer_generation.fetch_max(generation, Ordering::AcqRel);
+        if generation <= prev {
+            tracing::debug!(
+                generation,
+                prev,
+                "ice_agent: ignoring stale CandidateUpdate"
+            );
+            return None;
+        }
+
+        let reflexive = reflexive_addr
+            .as_deref()
+            .and_then(|s| s.parse().ok());
+        let local: Vec<SocketAddr> = local_addrs
+            .iter()
+            .filter_map(|s| s.parse().ok())
+            .collect();
+        let mapped = mapped_addr
+            .as_deref()
+            .and_then(|s| s.parse().ok());
+
+        tracing::info!(
+            generation,
+            reflexive = ?reflexive,
+            mapped = ?mapped,
+            local_count = local.len(),
+            "ice_agent: applied peer candidate update"
+        );
+
+        Some(PeerCandidates {
+            reflexive,
+            local,
+            mapped,
+        })
+    }
+
+    /// Get the current generation counter.
+    pub fn generation(&self) -> u32 {
+        self.generation.load(Ordering::Relaxed)
+    }
+}
+
+// ── Tests ──────────────────────────────────────────────────────────
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn apply_peer_update_rejects_stale() {
+        let agent = IceAgent::new("test-call".into(), IceAgentConfig::default());
+
+        // First update (gen=1) should succeed.
+        let update1 = SignalMessage::CandidateUpdate {
+            call_id: "test-call".into(),
+            reflexive_addr: Some("203.0.113.5:4433".into()),
+            local_addrs: vec!["192.168.1.10:4433".into()],
+            mapped_addr: None,
+            generation: 1,
+        };
+        let result = agent.apply_peer_update(&update1);
+        assert!(result.is_some());
+        let candidates = result.unwrap();
+        assert_eq!(
+            candidates.reflexive,
+            Some("203.0.113.5:4433".parse().unwrap())
+        );
+        assert_eq!(candidates.local.len(), 1);
+
+        // Same generation (gen=1) should be rejected.
+        let update1b = SignalMessage::CandidateUpdate {
+            call_id: "test-call".into(),
+            reflexive_addr: Some("198.51.100.9:4433".into()),
+            local_addrs: vec![],
+            mapped_addr: None,
+            generation: 1,
+        };
+        assert!(agent.apply_peer_update(&update1b).is_none());
+
+        // Older generation (gen=0) should be rejected.
+        let update0 = SignalMessage::CandidateUpdate {
+            call_id: "test-call".into(),
+            reflexive_addr: Some("10.0.0.1:4433".into()),
+            local_addrs: vec![],
+            mapped_addr: None,
+            generation: 0,
+        };
+        assert!(agent.apply_peer_update(&update0).is_none());
+
+        // Newer generation (gen=2) should succeed.
+        let update2 = SignalMessage::CandidateUpdate {
+            call_id: "test-call".into(),
+            reflexive_addr: Some("198.51.100.9:5555".into()),
+            local_addrs: vec![],
+            mapped_addr: Some("203.0.113.5:12345".into()),
+            generation: 2,
+        };
+        let result = agent.apply_peer_update(&update2);
+        assert!(result.is_some());
+        let candidates = result.unwrap();
+        assert_eq!(
+            candidates.reflexive,
+            Some("198.51.100.9:5555".parse().unwrap())
+        );
+        assert_eq!(
+            candidates.mapped,
+            Some("203.0.113.5:12345".parse().unwrap())
+        );
+    }
+
+    #[test]
+    fn apply_wrong_signal_returns_none() {
+        let agent = IceAgent::new("test-call".into(), IceAgentConfig::default());
+        let wrong = SignalMessage::Reflect;
+        assert!(agent.apply_peer_update(&wrong).is_none());
+    }
+
+    #[test]
+    fn generation_increments() {
+        let agent = IceAgent::new("test".into(), IceAgentConfig::default());
+        assert_eq!(agent.generation(), 0);
+        // Simulate what gather() does internally
+        let g1 = agent.generation.fetch_add(1, Ordering::Relaxed);
+        assert_eq!(g1, 0);
+        assert_eq!(agent.generation(), 1);
+        let g2 = agent.generation.fetch_add(1, Ordering::Relaxed);
+        assert_eq!(g2, 1);
+        assert_eq!(agent.generation(), 2);
+    }
+
+    #[test]
+    fn apply_peer_update_parses_all_fields() {
+        let agent = IceAgent::new("test-call".into(), IceAgentConfig::default());
+
+        let update = SignalMessage::CandidateUpdate {
+            call_id: "test-call".into(),
+            reflexive_addr: Some("203.0.113.5:4433".into()),
+            local_addrs: vec![
+                "192.168.1.10:4433".into(),
+                "10.0.0.5:4433".into(),
+            ],
+            mapped_addr: Some("198.51.100.42:12345".into()),
+            generation: 1,
+        };
+
+        let candidates = agent.apply_peer_update(&update).unwrap();
+        assert_eq!(
+            candidates.reflexive,
+            Some("203.0.113.5:4433".parse().unwrap())
+        );
+        assert_eq!(candidates.local.len(), 2);
+        assert_eq!(
+            candidates.local[0],
+            "192.168.1.10:4433".parse::<SocketAddr>().unwrap()
+        );
+        assert_eq!(
+            candidates.mapped,
+            Some("198.51.100.42:12345".parse().unwrap())
+        );
+    }
+
+    #[test]
+    fn apply_peer_update_handles_empty_fields() {
+        let agent = IceAgent::new("test".into(), IceAgentConfig::default());
+
+        let update = SignalMessage::CandidateUpdate {
+            call_id: "test".into(),
+            reflexive_addr: None,
+            local_addrs: vec![],
+            mapped_addr: None,
+            generation: 1,
+        };
+
+        let candidates = agent.apply_peer_update(&update).unwrap();
+        assert!(candidates.reflexive.is_none());
+        assert!(candidates.local.is_empty());
+        assert!(candidates.mapped.is_none());
+    }
+
+    #[test]
+    fn apply_peer_update_skips_unparseable_addrs() {
+        let agent = IceAgent::new("test".into(), IceAgentConfig::default());
+
+        let update = SignalMessage::CandidateUpdate {
+            call_id: "test".into(),
+            reflexive_addr: Some("not-an-addr".into()),
+            local_addrs: vec![
+                "192.168.1.10:4433".into(),
+                "garbage".into(),
+                "10.0.0.5:4433".into(),
+            ],
+            mapped_addr: Some("also-bad".into()),
+            generation: 1,
+        };
+
+        let candidates = agent.apply_peer_update(&update).unwrap();
+        assert!(candidates.reflexive.is_none()); // unparseable
+        assert_eq!(candidates.local.len(), 2); // garbage filtered
+        assert!(candidates.mapped.is_none()); // unparseable
+    }
+
+    #[test]
+    fn default_config_values() {
+        let cfg = IceAgentConfig::default();
+        assert!(cfg.enable_portmap);
+        assert!(cfg.gather_timeout.as_secs() > 0);
+        assert!(!cfg.stun_config.servers.is_empty());
+        assert_eq!(cfg.local_v4_port, 0);
+        assert!(cfg.local_v6_port.is_none());
+    }
+
+    #[tokio::test]
+    async fn gather_returns_candidates_even_with_no_stun() {
+        // With default config (port 0 = no portmap, STUN will timeout
+        // quickly on loopback), gather should still return host candidates.
+        let agent = IceAgent::new("test".into(), IceAgentConfig {
+            stun_config: stun::StunConfig {
+                servers: vec![], // no servers = quick failure
+                timeout: Duration::from_millis(100),
+            },
+            enable_portmap: false,
+            gather_timeout: Duration::from_millis(200),
+            local_v4_port: 12345,
+            local_v6_port: None,
+        });
+
+        let candidates = agent.gather().await;
+        assert_eq!(candidates.generation, 0);
+        // Reflexive should be None (no STUN servers)
+        assert!(candidates.reflexive.is_none());
+        // Mapped should be None (portmap disabled)
+        assert!(candidates.mapped.is_none());
+        // Local candidates depend on the machine's interfaces
+        // but gather() should not panic.
+    }
+
+    #[tokio::test]
+    async fn re_gather_produces_signal_message() {
+        let agent = IceAgent::new("call-42".into(), IceAgentConfig {
+            stun_config: stun::StunConfig {
+                servers: vec![],
+                timeout: Duration::from_millis(50),
+            },
+            enable_portmap: false,
+            gather_timeout: Duration::from_millis(100),
+            local_v4_port: 4433,
+            local_v6_port: None,
+        });
+
+        let (candidates, signal) = agent.re_gather().await;
+        assert_eq!(candidates.generation, 0);
+
+        match signal {
+            SignalMessage::CandidateUpdate {
+                call_id,
+                generation,
+                ..
+            } => {
+                assert_eq!(call_id, "call-42");
+                assert_eq!(generation, 0);
+            }
+            _ => panic!("expected CandidateUpdate"),
+        }
+
+        // Second re_gather increments generation
+        let (candidates2, signal2) = agent.re_gather().await;
+        assert_eq!(candidates2.generation, 1);
+        match signal2 {
+            SignalMessage::CandidateUpdate { generation, .. } => {
+                assert_eq!(generation, 1);
+            }
+            _ => panic!("expected CandidateUpdate"),
+        }
+    }
+}

crates/wzp-client/src/lib.rs

@@ -8,16 +8,83 @@
 
 #[cfg(feature = "audio")]
 pub mod audio_io;
+#[cfg(feature = "audio")]
+pub mod audio_ring;
+// VoiceProcessingIO is an Apple Core Audio API — only compile the module
+// when the `vpio` feature is on AND we're targeting macOS. Enabling the
+// feature on Windows/Linux was previously silently broken.
+#[cfg(all(feature = "vpio", target_os = "macos"))]
+pub mod audio_vpio;
+// WASAPI-direct capture with Windows's OS-level AEC (AudioCategory_Communications).
+// Only compiled when `windows-aec` feature is on AND target is Windows. The
+// `windows` dependency is itself gated to Windows in Cargo.toml, so enabling
+// this feature on non-Windows targets is a no-op.
+#[cfg(all(feature = "windows-aec", target_os = "windows"))]
+pub mod audio_wasapi;
+// WebRTC AEC3 (Audio Processing Module) wrapper around CPAL capture + playback
+// on Linux. Only compiled when `linux-aec` feature is on AND target is Linux.
+// The webrtc-audio-processing dep is itself gated to Linux in Cargo.toml.
+#[cfg(all(feature = "linux-aec", target_os = "linux"))]
+pub mod audio_linux_aec;
 pub mod bench;
 pub mod call;
 pub mod drift_test;
 pub mod echo_test;
 pub mod featherchat;
 pub mod handshake;
+pub mod dual_path;
 pub mod metrics;
+pub mod birthday;
+pub mod ice_agent;
+pub mod netcheck;
+pub mod portmap;
+pub mod reflect;
+pub mod relay_map;
+pub mod stun;
 pub mod sweep;
 
-#[cfg(feature = "audio")]
-pub use audio_io::{AudioCapture, AudioPlayback};
+// AudioPlayback: three possible backends depending on feature flags.
+//   1. Default CPAL (`audio_io::AudioPlayback`) — baseline on every platform.
+//   2. Linux AEC (`audio_linux_aec::LinuxAecPlayback`) — CPAL + WebRTC APM
+//      render-side tee, so echo from speakers gets cancelled from the mic.
+//
+// On macOS and Windows we always use the default CPAL playback because:
+//   - macOS: VoiceProcessingIO handles AEC at the capture side (Apple's
+//     native hardware AEC uses its own reference signal handling).
+//   - Windows: WASAPI AudioCategory_Communications AEC uses the system
+//     render mix as reference — no per-process plumbing needed.
+//
+// Linux is the only platform where the in-app approach is necessary, so
+// the AEC playback path is gated to target_os = "linux".
+
+#[cfg(all(
+    feature = "audio",
+    any(not(feature = "linux-aec"), not(target_os = "linux"))
+))]
+pub use audio_io::AudioPlayback;
+
+#[cfg(all(feature = "linux-aec", target_os = "linux"))]
+pub use audio_linux_aec::LinuxAecPlayback as AudioPlayback;
+
+// AudioCapture: three possible backends depending on feature flags.
+//   1. Default CPAL (`audio_io::AudioCapture`) — baseline on every platform.
+//   2. Windows AEC (`audio_wasapi::WasapiAudioCapture`) — direct WASAPI
+//      with AudioCategory_Communications, OS APO chain does AEC.
+//   3. Linux AEC (`audio_linux_aec::LinuxAecCapture`) — CPAL + WebRTC APM
+//      capture-side echo cancellation using the playback tee as reference.
+// All three expose the same public API (`start`, `ring`, `stop`, `Drop`).
+
+#[cfg(all(
+    feature = "audio",
+    any(not(feature = "windows-aec"), not(target_os = "windows")),
+    any(not(feature = "linux-aec"), not(target_os = "linux"))
+))]
+pub use audio_io::AudioCapture;
+
+#[cfg(all(feature = "windows-aec", target_os = "windows"))]
+pub use audio_wasapi::WasapiAudioCapture as AudioCapture;
+
+#[cfg(all(feature = "linux-aec", target_os = "linux"))]
+pub use audio_linux_aec::LinuxAecCapture as AudioCapture;
 pub use call::{CallConfig, CallDecoder, CallEncoder};
 pub use handshake::perform_handshake;

crates/wzp-client/src/netcheck.rs

@@ -0,0 +1,524 @@
+//! Phase 8 (Tailscale-inspired): Comprehensive network diagnostic.
+//!
+//! Probes STUN servers, relay infrastructure, port mapping
+//! capabilities, IPv6 reachability, and NAT hairpinning in parallel
+//! to produce a `NetcheckReport` that captures the client's network
+//! environment at a point in time.
+//!
+//! Used for:
+//! - Troubleshooting connectivity issues
+//! - Automatic relay selection (Phase 5)
+//! - Pre-call NAT assessment
+//! - Quality prediction
+
+use std::net::SocketAddr;
+use std::time::{Duration, Instant};
+
+use serde::Serialize;
+
+use crate::portmap::{self, PortMapProtocol};
+use crate::reflect::{self, NatType};
+use crate::stun::{self, StunConfig};
+
+/// Complete network diagnostic report.
+#[derive(Debug, Clone, Serialize)]
+pub struct NetcheckReport {
+    /// NAT type classification (from combined STUN + relay probes).
+    pub nat_type: NatType,
+    /// Server-reflexive address (consensus from probes).
+    pub reflexive_addr: Option<String>,
+    /// Whether IPv4 connectivity is available.
+    pub ipv4_reachable: bool,
+    /// Whether IPv6 connectivity is available.
+    pub ipv6_reachable: bool,
+    /// Whether the NAT supports hairpinning (loopback to own
+    /// reflexive address).
+    pub hairpin_works: Option<bool>,
+    /// Which port mapping protocol is available (if any).
+    pub port_mapping: Option<PortMapProtocol>,
+    /// Per-relay latency measurements.
+    pub relay_latencies: Vec<RelayLatency>,
+    /// Preferred relay (lowest latency).
+    pub preferred_relay: Option<String>,
+    /// STUN latency to first responding server (ms).
+    pub stun_latency_ms: Option<u32>,
+    /// Whether UPnP is available on the gateway.
+    pub upnp_available: bool,
+    /// Whether PCP is available on the gateway.
+    pub pcp_available: bool,
+    /// Whether NAT-PMP is available on the gateway.
+    pub nat_pmp_available: bool,
+    /// Default gateway address.
+    pub gateway: Option<String>,
+    /// Total time taken for the diagnostic (ms).
+    pub duration_ms: u32,
+    /// Individual STUN probe results.
+    pub stun_probes: Vec<reflect::NatProbeResult>,
+    /// NAT port allocation pattern (sequential vs random).
+    pub port_allocation: Option<stun::PortAllocation>,
+}
+
+/// Latency to a specific relay.
+#[derive(Debug, Clone, Serialize)]
+pub struct RelayLatency {
+    pub name: String,
+    pub addr: String,
+    pub rtt_ms: Option<u32>,
+    pub error: Option<String>,
+}
+
+/// Configuration for the netcheck run.
+#[derive(Debug, Clone)]
+pub struct NetcheckConfig {
+    /// STUN servers to probe.
+    pub stun_config: StunConfig,
+    /// Relay servers to probe (name, address pairs).
+    pub relays: Vec<(String, SocketAddr)>,
+    /// Per-probe timeout.
+    pub timeout: Duration,
+    /// Whether to test port mapping.
+    pub test_portmap: bool,
+    /// Whether to test IPv6.
+    pub test_ipv6: bool,
+    /// Local port for port mapping test (0 = skip).
+    pub local_port: u16,
+}
+
+impl Default for NetcheckConfig {
+    fn default() -> Self {
+        Self {
+            stun_config: StunConfig::default(),
+            relays: Vec::new(),
+            timeout: Duration::from_secs(5),
+            test_portmap: true,
+            test_ipv6: true,
+            local_port: 0,
+        }
+    }
+}
+
+/// Run a comprehensive network diagnostic.
+///
+/// Probes run in parallel for speed — the total time is bounded
+/// by the slowest individual probe, not the sum.
+pub async fn run_netcheck(config: &NetcheckConfig) -> NetcheckReport {
+    let start = Instant::now();
+
+    // Run all probes in parallel.
+    let stun_fut = stun::probe_stun_servers(&config.stun_config);
+    let relay_fut = probe_relays(&config.relays, config.timeout);
+    let portmap_fut = probe_portmap(config.test_portmap, config.local_port);
+    let gateway_fut = portmap::default_gateway();
+    let ipv6_fut = test_ipv6(config.test_ipv6, config.timeout);
+    let port_alloc_fut = stun::detect_port_allocation(&config.stun_config);
+
+    let (stun_probes, relay_latencies, portmap_result, gateway_result, ipv6_reachable, port_alloc_result) =
+        tokio::join!(stun_fut, relay_fut, portmap_fut, gateway_result_fut(gateway_fut), ipv6_fut, port_alloc_fut);
+
+    // Classify NAT from STUN probes.
+    let (nat_type, consensus_addr) = reflect::classify_nat(&stun_probes);
+
+    // Determine STUN latency (first successful probe).
+    let stun_latency_ms = stun_probes
+        .iter()
+        .filter_map(|p| p.latency_ms)
+        .min();
+
+    // IPv4 reachable if any STUN probe succeeded.
+    let ipv4_reachable = stun_probes
+        .iter()
+        .any(|p| p.observed_addr.is_some());
+
+    // Preferred relay = lowest RTT.
+    let preferred_relay = relay_latencies
+        .iter()
+        .filter_map(|r| r.rtt_ms.map(|rtt| (r.name.clone(), rtt)))
+        .min_by_key(|(_, rtt)| *rtt)
+        .map(|(name, _)| name);
+
+    // Port mapping availability.
+    let (port_mapping, nat_pmp_available, pcp_available, upnp_available) = match portmap_result {
+        Some(mapping) => {
+            let proto = mapping.protocol;
+            (
+                Some(proto),
+                proto == PortMapProtocol::NatPmp,
+                proto == PortMapProtocol::Pcp,
+                proto == PortMapProtocol::UPnP,
+            )
+        }
+        None => (None, false, false, false),
+    };
+
+    let gateway = match gateway_result {
+        Ok(gw) => Some(gw.to_string()),
+        Err(_) => None,
+    };
+
+    NetcheckReport {
+        nat_type,
+        reflexive_addr: consensus_addr,
+        ipv4_reachable,
+        ipv6_reachable,
+        hairpin_works: None, // TODO: implement hairpin test
+        port_mapping,
+        relay_latencies,
+        preferred_relay,
+        stun_latency_ms,
+        upnp_available,
+        pcp_available,
+        nat_pmp_available,
+        gateway,
+        duration_ms: start.elapsed().as_millis() as u32,
+        stun_probes,
+        port_allocation: Some(port_alloc_result.allocation),
+    }
+}
+
+/// Probe relay latencies via reflect.
+async fn probe_relays(
+    relays: &[(String, SocketAddr)],
+    timeout: Duration,
+) -> Vec<RelayLatency> {
+    if relays.is_empty() {
+        return Vec::new();
+    }
+
+    let timeout_ms = timeout.as_millis() as u64;
+    let mut set = tokio::task::JoinSet::new();
+
+    for (name, addr) in relays {
+        let name = name.clone();
+        let addr = *addr;
+        set.spawn(async move {
+            let start = Instant::now();
+            match reflect::probe_reflect_addr(addr, timeout_ms, None).await {
+                Ok((_observed, _latency)) => RelayLatency {
+                    name,
+                    addr: addr.to_string(),
+                    rtt_ms: Some(start.elapsed().as_millis() as u32),
+                    error: None,
+                },
+                Err(e) => RelayLatency {
+                    name,
+                    addr: addr.to_string(),
+                    rtt_ms: None,
+                    error: Some(e),
+                },
+            }
+        });
+    }
+
+    let mut results = Vec::with_capacity(relays.len());
+    while let Some(join_result) = set.join_next().await {
+        match join_result {
+            Ok(r) => results.push(r),
+            Err(_) => {}
+        }
+    }
+
+    // Sort by RTT (lowest first).
+    results.sort_by_key(|r| r.rtt_ms.unwrap_or(u32::MAX));
+    results
+}
+
+/// Attempt port mapping and return the mapping if successful.
+async fn probe_portmap(
+    enabled: bool,
+    local_port: u16,
+) -> Option<portmap::PortMapping> {
+    if !enabled || local_port == 0 {
+        return None;
+    }
+    portmap::acquire_port_mapping(local_port, None).await.ok()
+}
+
+/// Wrap the gateway future to handle the Result.
+async fn gateway_result_fut(
+    fut: impl std::future::Future<Output = Result<std::net::Ipv4Addr, portmap::PortMapError>>,
+) -> Result<std::net::Ipv4Addr, portmap::PortMapError> {
+    fut.await
+}
+
+/// Test IPv6 connectivity by attempting to bind and send on an IPv6 socket.
+async fn test_ipv6(enabled: bool, timeout: Duration) -> bool {
+    if !enabled {
+        return false;
+    }
+
+    // Try to resolve and connect to an IPv6 STUN server.
+    let result = tokio::time::timeout(timeout, async {
+        let sock = tokio::net::UdpSocket::bind("[::]:0").await.ok()?;
+        // Try Google's IPv6 STUN — if DNS resolves to an AAAA record
+        // and we can send a packet, IPv6 is working.
+        let addr = stun::resolve_stun_server("stun.l.google.com:19302").await.ok()?;
+        if addr.is_ipv6() {
+            sock.send_to(&[0u8; 1], addr).await.ok()?;
+            Some(true)
+        } else {
+            // Server resolved to IPv4 — try binding to [::] at least
+            Some(false)
+        }
+    })
+    .await;
+
+    match result {
+        Ok(Some(true)) => true,
+        _ => {
+            // Fallback: can we at least bind an IPv6 socket?
+            tokio::net::UdpSocket::bind("[::]:0").await.is_ok()
+        }
+    }
+}
+
+/// Format a netcheck report as a human-readable string.
+pub fn format_report(report: &NetcheckReport) -> String {
+    let mut out = String::new();
+
+    out.push_str(&format!("=== WarzonePhone Netcheck ===\n\n"));
+    out.push_str(&format!(
+        "NAT Type:       {:?}\n",
+        report.nat_type
+    ));
+    out.push_str(&format!(
+        "Reflexive Addr: {}\n",
+        report.reflexive_addr.as_deref().unwrap_or("(unknown)")
+    ));
+    out.push_str(&format!(
+        "IPv4:           {}\n",
+        if report.ipv4_reachable { "yes" } else { "no" }
+    ));
+    out.push_str(&format!(
+        "IPv6:           {}\n",
+        if report.ipv6_reachable { "yes" } else { "no" }
+    ));
+    out.push_str(&format!(
+        "Gateway:        {}\n",
+        report.gateway.as_deref().unwrap_or("(unknown)")
+    ));
+
+    if let Some(ref alloc) = report.port_allocation {
+        out.push_str(&format!(
+            "Port Alloc:     {alloc}\n"
+        ));
+    }
+
+    out.push_str(&format!("\n--- Port Mapping ---\n"));
+    out.push_str(&format!(
+        "NAT-PMP: {}  PCP: {}  UPnP: {}\n",
+        if report.nat_pmp_available { "yes" } else { "no" },
+        if report.pcp_available { "yes" } else { "no" },
+        if report.upnp_available { "yes" } else { "no" },
+    ));
+    if let Some(proto) = &report.port_mapping {
+        out.push_str(&format!("Active mapping: {:?}\n", proto));
+    }
+
+    if !report.stun_probes.is_empty() {
+        out.push_str(&format!("\n--- STUN Probes ---\n"));
+        for p in &report.stun_probes {
+            out.push_str(&format!(
+                "  {} → {} ({}ms){}\n",
+                p.relay_name,
+                p.observed_addr.as_deref().unwrap_or("failed"),
+                p.latency_ms.map(|ms| ms.to_string()).unwrap_or_else(|| "-".into()),
+                p.error.as_ref().map(|e| format!(" [{e}]")).unwrap_or_default(),
+            ));
+        }
+    }
+
+    if !report.relay_latencies.is_empty() {
+        out.push_str(&format!("\n--- Relay Latencies ---\n"));
+        for r in &report.relay_latencies {
+            out.push_str(&format!(
+                "  {} ({}) → {}ms{}\n",
+                r.name,
+                r.addr,
+                r.rtt_ms.map(|ms| ms.to_string()).unwrap_or_else(|| "-".into()),
+                r.error.as_ref().map(|e| format!(" [{e}]")).unwrap_or_default(),
+            ));
+        }
+        if let Some(ref pref) = report.preferred_relay {
+            out.push_str(&format!("  Preferred: {pref}\n"));
+        }
+    }
+
+    out.push_str(&format!("\nCompleted in {}ms\n", report.duration_ms));
+    out
+}
+
+// ── Tests ──────────────────────────────────────────────────────────
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn default_config_has_stun_servers() {
+        let config = NetcheckConfig::default();
+        assert!(!config.stun_config.servers.is_empty());
+    }
+
+    #[test]
+    fn format_report_produces_output() {
+        let report = NetcheckReport {
+            nat_type: NatType::Cone,
+            reflexive_addr: Some("203.0.113.5:4433".into()),
+            ipv4_reachable: true,
+            ipv6_reachable: false,
+            hairpin_works: None,
+            port_mapping: None,
+            relay_latencies: vec![RelayLatency {
+                name: "relay-1".into(),
+                addr: "10.0.0.1:4433".into(),
+                rtt_ms: Some(25),
+                error: None,
+            }],
+            preferred_relay: Some("relay-1".into()),
+            stun_latency_ms: Some(15),
+            upnp_available: false,
+            pcp_available: false,
+            nat_pmp_available: false,
+            gateway: Some("192.168.1.1".into()),
+            duration_ms: 1500,
+            stun_probes: vec![],
+            port_allocation: None,
+        };
+
+        let text = format_report(&report);
+        assert!(text.contains("Cone"));
+        assert!(text.contains("203.0.113.5:4433"));
+        assert!(text.contains("relay-1"));
+        assert!(text.contains("1500ms"));
+    }
+
+    #[test]
+    fn report_serializes_to_json() {
+        let report = NetcheckReport {
+            nat_type: NatType::Cone,
+            reflexive_addr: Some("203.0.113.5:4433".into()),
+            ipv4_reachable: true,
+            ipv6_reachable: false,
+            hairpin_works: None,
+            port_mapping: Some(PortMapProtocol::NatPmp),
+            relay_latencies: vec![],
+            preferred_relay: None,
+            stun_latency_ms: Some(25),
+            upnp_available: false,
+            pcp_available: false,
+            nat_pmp_available: true,
+            gateway: Some("192.168.1.1".into()),
+            duration_ms: 500,
+            stun_probes: vec![],
+            port_allocation: Some(stun::PortAllocation::Sequential { delta: 1 }),
+        };
+        let json = serde_json::to_string(&report).unwrap();
+        assert!(json.contains("Cone"));
+        assert!(json.contains("203.0.113.5:4433"));
+        assert!(json.contains("NatPmp"));
+
+        // Roundtrip
+        let decoded: serde_json::Value = serde_json::from_str(&json).unwrap();
+        assert_eq!(decoded["ipv4_reachable"], true);
+        assert_eq!(decoded["ipv6_reachable"], false);
+        assert_eq!(decoded["stun_latency_ms"], 25);
+    }
+
+    #[test]
+    fn relay_latency_serializes() {
+        let lat = RelayLatency {
+            name: "eu-west".into(),
+            addr: "10.0.0.1:4433".into(),
+            rtt_ms: Some(42),
+            error: None,
+        };
+        let json = serde_json::to_string(&lat).unwrap();
+        assert!(json.contains("eu-west"));
+        assert!(json.contains("42"));
+    }
+
+    #[test]
+    fn format_report_empty_relays() {
+        let report = NetcheckReport {
+            nat_type: NatType::Unknown,
+            reflexive_addr: None,
+            ipv4_reachable: false,
+            ipv6_reachable: false,
+            hairpin_works: None,
+            port_mapping: None,
+            relay_latencies: vec![],
+            preferred_relay: None,
+            stun_latency_ms: None,
+            upnp_available: false,
+            pcp_available: false,
+            nat_pmp_available: false,
+            gateway: None,
+            duration_ms: 100,
+            stun_probes: vec![],
+            port_allocation: None,
+        };
+        let text = format_report(&report);
+        assert!(text.contains("Unknown"));
+        assert!(text.contains("(unknown)")); // reflexive addr
+        assert!(text.contains("100ms"));
+    }
+
+    #[test]
+    fn format_report_with_stun_probes() {
+        let report = NetcheckReport {
+            nat_type: NatType::SymmetricPort,
+            reflexive_addr: None,
+            ipv4_reachable: true,
+            ipv6_reachable: true,
+            hairpin_works: Some(false),
+            port_mapping: Some(PortMapProtocol::UPnP),
+            relay_latencies: vec![
+                RelayLatency {
+                    name: "us-east".into(),
+                    addr: "10.0.0.1:4433".into(),
+                    rtt_ms: Some(15),
+                    error: None,
+                },
+                RelayLatency {
+                    name: "eu-west".into(),
+                    addr: "10.0.0.2:4433".into(),
+                    rtt_ms: None,
+                    error: Some("timeout".into()),
+                },
+            ],
+            preferred_relay: Some("us-east".into()),
+            stun_latency_ms: Some(20),
+            upnp_available: true,
+            pcp_available: false,
+            nat_pmp_available: false,
+            gateway: Some("192.168.0.1".into()),
+            duration_ms: 3000,
+            stun_probes: vec![reflect::NatProbeResult {
+                relay_name: "stun:google".into(),
+                relay_addr: "74.125.250.129:19302".into(),
+                observed_addr: Some("203.0.113.5:12345".into()),
+                latency_ms: Some(20),
+                error: None,
+            }],
+            port_allocation: Some(stun::PortAllocation::Random),
+        };
+        let text = format_report(&report);
+        assert!(text.contains("SymmetricPort"));
+        assert!(text.contains("us-east"));
+        assert!(text.contains("eu-west"));
+        assert!(text.contains("Preferred: us-east"));
+        assert!(text.contains("UPnP: yes"));
+        assert!(text.contains("stun:google"));
+        assert!(text.contains("3000ms"));
+    }
+
+    /// Integration test: run actual netcheck (requires network).
+    #[tokio::test]
+    #[ignore]
+    async fn integration_netcheck() {
+        let config = NetcheckConfig::default();
+        let report = run_netcheck(&config).await;
+        println!("{}", format_report(&report));
+        assert!(report.duration_ms > 0);
+    }
+}

crates/wzp-client/src/reflect.rs

@@ -0,0 +1,713 @@
+//! Multi-relay NAT reflection ("STUN for QUIC" — Phase 2).
+//!
+//! Phase 1 (`SignalMessage::Reflect` / `ReflectResponse`) lets a
+//! client ask a single relay "what source address do you see for
+//! me?". Phase 2 queries N relays in parallel and classifies the
+//! results into a NAT type so the future P2P hole-punching path
+//! can decide whether a direct QUIC handshake is viable:
+//!
+//! - All relays return the same `(ip, port)` → **Cone NAT**.
+//!   Endpoint-independent mapping, P2P hole-punching viable,
+//!   `consensus_addr` is the one address to advertise.
+//! - Same ip, different ports → **Symmetric port-dependent NAT**.
+//!   The mapping changes per destination, so the advertised addr
+//!   wouldn't match what a peer actually sees; fall back to
+//!   relay-mediated path.
+//! - Different ips → multi-homed / anycast / broken DNS, treat as
+//!   `Multiple` and do not attempt P2P.
+//! - 0 or 1 successful probes → `Unknown`, not enough data.
+//!
+//! A probe is a throwaway QUIC signal connection: open endpoint,
+//! connect, RegisterPresence (with a zero identity — the relay
+//! accepts this exactly like the main signaling path does), send
+//! Reflect, read ReflectResponse, close. Each probe gets its own
+//! ephemeral quinn::Endpoint so the OS assigns a fresh source port
+//! per relay — if we shared one endpoint across probes, a
+//! symmetric NAT in front of the client would map every probe to
+//! the same port and we couldn't detect it.
+
+use std::net::SocketAddr;
+use std::time::{Duration, Instant};
+
+use serde::Serialize;
+use wzp_proto::{MediaTransport, SignalMessage};
+use wzp_transport::{client_config, create_endpoint, QuinnTransport};
+
+/// Result of one probe against one relay. Always returned so the
+/// UI can render per-relay status even when some fail.
+#[derive(Debug, Clone, Serialize)]
+pub struct NatProbeResult {
+    pub relay_name: String,
+    pub relay_addr: String,
+    /// `Some` on successful probe, `None` on failure.
+    pub observed_addr: Option<String>,
+    /// End-to-end wall-clock from connect start to ReflectResponse
+    /// received, in milliseconds. `Some` only on success.
+    pub latency_ms: Option<u32>,
+    /// Human-readable error on failure.
+    pub error: Option<String>,
+}
+
+/// Aggregated classification over N `NatProbeResult`s.
+#[derive(Debug, Clone, Serialize)]
+pub struct NatDetection {
+    pub probes: Vec<NatProbeResult>,
+    pub nat_type: NatType,
+    /// When `nat_type == Cone`, the one address all probes agreed
+    /// on. `None` for every other case.
+    pub consensus_addr: Option<String>,
+}
+
+/// NAT classification. See module doc for semantics.
+#[derive(Debug, Clone, Copy, Serialize, PartialEq, Eq)]
+pub enum NatType {
+    Cone,
+    SymmetricPort,
+    Multiple,
+    Unknown,
+}
+
+/// Probe a single relay with a QUIC connection.
+///
+/// # Endpoint reuse (Phase 5 — Nebula-style architecture)
+///
+/// If `existing_endpoint` is `Some`, the probe uses that socket
+/// instead of creating a fresh one. This is the desired mode in
+/// production: a port-preserving NAT (MikroTik masquerade, most
+/// consumer routers) gives a **stable** external port for the
+/// one socket, so the reflex addr observed by ANY relay is the
+/// SAME addr and matches what a peer would see on a direct dial.
+/// Pass the signal endpoint here.
+///
+/// If `None`, creates a fresh one-shot endpoint. Kept for:
+/// - tests that spin up isolated probes
+/// - the "I'm not registered yet" case where there's no signal
+///   endpoint to reuse
+///
+/// NOTE on NAT-type detection: the pre-Phase-5 behavior of
+/// forcing a fresh endpoint per probe was wrong — it made every
+/// port-preserving NAT look symmetric because the classifier saw
+/// a different external port for each fresh source port. With
+/// one shared socket, the classifier reflects the REAL NAT
+/// behavior.
+pub async fn probe_reflect_addr(
+    relay: SocketAddr,
+    timeout_ms: u64,
+    existing_endpoint: Option<wzp_transport::Endpoint>,
+) -> Result<(SocketAddr, u32), String> {
+    // Install rustls provider idempotently — a second install on the
+    // same thread is a no-op.
+    let _ = rustls::crypto::ring::default_provider().install_default();
+
+    let endpoint = match existing_endpoint {
+        Some(ep) => ep,
+        None => {
+            let bind: SocketAddr = "0.0.0.0:0".parse().unwrap();
+            create_endpoint(bind, None).map_err(|e| format!("endpoint: {e}"))?
+        }
+    };
+
+    let start = Instant::now();
+    let probe = async {
+        // Open the signal connection.
+        let conn =
+            wzp_transport::connect(&endpoint, relay, "_signal", client_config())
+                .await
+                .map_err(|e| format!("connect: {e}"))?;
+        let transport = QuinnTransport::new(conn);
+
+        // The relay signal handler waits for a RegisterPresence
+        // before entering its main dispatch loop (see
+        // wzp-relay/src/main.rs). So a transient probe has to
+        // register with a zero identity first — the relay accepts
+        // the empty-signature form exactly as the main signaling
+        // path does in desktop/src-tauri/src/lib.rs register_signal.
+        transport
+            .send_signal(&SignalMessage::RegisterPresence {
+                identity_pub: [0u8; 32],
+                signature: vec![],
+                alias: None,
+            })
+            .await
+            .map_err(|e| format!("send RegisterPresence: {e}"))?;
+        // Drain the RegisterPresenceAck so the response to our
+        // Reflect doesn't land on an unexpected stream order.
+        match transport.recv_signal().await {
+            Ok(Some(SignalMessage::RegisterPresenceAck { success: true, .. })) => {}
+            Ok(Some(other)) => {
+                return Err(format!(
+                    "unexpected pre-reflect signal: {:?}",
+                    std::mem::discriminant(&other)
+                ));
+            }
+            Ok(None) => return Err("connection closed before RegisterPresenceAck".into()),
+            Err(e) => return Err(format!("recv RegisterPresenceAck: {e}")),
+        }
+
+        // Send Reflect and await response.
+        transport
+            .send_signal(&SignalMessage::Reflect)
+            .await
+            .map_err(|e| format!("send Reflect: {e}"))?;
+
+        match transport.recv_signal().await {
+            Ok(Some(SignalMessage::ReflectResponse { observed_addr })) => {
+                let parsed: SocketAddr = observed_addr
+                    .parse()
+                    .map_err(|e| format!("parse observed_addr {observed_addr:?}: {e}"))?;
+                let latency_ms = start.elapsed().as_millis() as u32;
+
+                // Clean close so the relay's per-connection cleanup
+                // runs promptly and we don't leak file descriptors.
+                let _ = transport.close().await;
+
+                Ok((parsed, latency_ms))
+            }
+            Ok(Some(other)) => Err(format!(
+                "expected ReflectResponse, got {:?}",
+                std::mem::discriminant(&other)
+            )),
+            Ok(None) => Err("connection closed before ReflectResponse".into()),
+            Err(e) => Err(format!("recv ReflectResponse: {e}")),
+        }
+    };
+
+    let out = tokio::time::timeout(Duration::from_millis(timeout_ms), probe)
+        .await
+        .map_err(|_| format!("probe timeout ({timeout_ms}ms)"))??;
+
+    // `endpoint` is a quinn::Endpoint clone — an Arc under the
+    // hood. Letting it drop at end-of-scope is correct whether it
+    // was fresh (last ref → socket closes) or shared (ref count
+    // decrements, socket stays alive for the signal loop).
+    Ok(out)
+}
+
+/// Detect the client's NAT type by probing N relays in parallel and
+/// classifying the returned addresses. Never errors — failing
+/// probes surface via `NatProbeResult.error`; aggregate is always
+/// returned.
+///
+/// # Endpoint reuse (Phase 5)
+///
+/// If `shared_endpoint` is `Some`, every probe reuses it. This is
+/// the PRODUCTION behavior: all probes source from the same UDP
+/// port, so port-preserving NATs map them to the same external
+/// port, and the classifier reflects the real NAT type. Pass the
+/// signal endpoint.
+///
+/// If `None`, each probe creates its own fresh endpoint — useful
+/// in tests that don't have a signal endpoint, but produces
+/// spurious `SymmetricPort` classifications against NATs that
+/// would otherwise look cone-like.
+pub async fn detect_nat_type(
+    relays: Vec<(String, SocketAddr)>,
+    timeout_ms: u64,
+    shared_endpoint: Option<wzp_transport::Endpoint>,
+) -> NatDetection {
+    // Parallel probes via tokio::task::JoinSet so the wall-clock is
+    // bounded by the slowest probe, not the sum. JoinSet keeps the
+    // dep surface at just tokio — we already depend on it.
+    let mut set = tokio::task::JoinSet::new();
+    for (name, addr) in relays {
+        let ep = shared_endpoint.clone();
+        set.spawn(async move {
+            let result = probe_reflect_addr(addr, timeout_ms, ep).await;
+            (name, addr, result)
+        });
+    }
+
+    let mut probes = Vec::new();
+    while let Some(join_result) = set.join_next().await {
+        let (name, addr, result) = match join_result {
+            Ok(tuple) => tuple,
+            // Task panicked — surface as a synthetic failed probe so
+            // the aggregate still returns a reasonable shape. This
+            // shouldn't happen but we don't want one bad probe to
+            // poison the whole detection.
+            Err(join_err) => {
+                probes.push(NatProbeResult {
+                    relay_name: "<panicked>".into(),
+                    relay_addr: "unknown".into(),
+                    observed_addr: None,
+                    latency_ms: None,
+                    error: Some(format!("probe task panicked: {join_err}")),
+                });
+                continue;
+            }
+        };
+        probes.push(match result {
+            Ok((observed, latency_ms)) => NatProbeResult {
+                relay_name: name,
+                relay_addr: addr.to_string(),
+                observed_addr: Some(observed.to_string()),
+                latency_ms: Some(latency_ms),
+                error: None,
+            },
+            Err(e) => NatProbeResult {
+                relay_name: name,
+                relay_addr: addr.to_string(),
+                observed_addr: None,
+                latency_ms: None,
+                error: Some(e),
+            },
+        });
+    }
+
+    let (nat_type, consensus_addr) = classify_nat(&probes);
+    NatDetection {
+        probes,
+        nat_type,
+        consensus_addr,
+    }
+}
+
+/// Enumerate LAN-local host candidates this client is reachable
+/// on, paired with the given port (typically the signal
+/// endpoint's bound port so that incoming dials land on the same
+/// socket the advertised reflex addr points to).
+///
+/// Gathers BOTH IPv4 and IPv6 candidates:
+///
+/// - **IPv4**: RFC1918 private ranges (10/8, 172.16/12, 192.168/16)
+///   and CGNAT shared-transition (100.64/10). Public IPv4 is
+///   skipped because the reflex-addr path already covers it.
+///   Loopback and link-local (169.254/16) are skipped.
+///
+/// - **IPv6**: ALL global-unicast addresses (2000::/3 — the real
+///   routable IPv6 space) AND unique-local (fc00::/7). These
+///   are directly dialable from a peer on the same LAN, and on
+///   true dual-stack LANs (which most consumer ISPs now provide,
+///   including Starlink) IPv6 often gives a direct path even
+///   when IPv4 can't hairpin. Loopback (::1), unspecified (::),
+///   and link-local (fe80::/10) are skipped — link-local would
+///   require a scope ID to be useful and is basically never
+///   reachable across interface boundaries.
+///
+/// The port must come from the caller — typically
+/// `signal_endpoint.local_addr()?.port()`, so that the peer's
+/// dials to these addresses land on the same socket that's
+/// already listening (Phase 5 shared-endpoint architecture).
+///
+/// Safe to call from any thread; no I/O, no async. The `if-addrs`
+/// crate reads the kernel's interface table via a single
+/// getifaddrs(3) syscall.
+pub fn local_host_candidates(v4_port: u16, v6_port: Option<u16>) -> Vec<SocketAddr> {
+    let Ok(ifaces) = if_addrs::get_if_addrs() else {
+        return Vec::new();
+    };
+    let mut out = Vec::new();
+    for iface in ifaces {
+        if iface.is_loopback() {
+            continue;
+        }
+        match iface.ip() {
+            std::net::IpAddr::V4(v4) => {
+                if v4.is_link_local() {
+                    continue;
+                }
+                // Keep RFC1918 private ranges and CGNAT — those
+                // are the LAN-dialable addrs we actually want.
+                // Skip public v4 because the reflex addr already
+                // covers that path.
+                if v4.is_private() {
+                    out.push(SocketAddr::new(std::net::IpAddr::V4(v4), v4_port));
+                } else if v4.octets()[0] == 100 && (v4.octets()[1] & 0xc0) == 0x40 {
+                    // 100.64/10 CGNAT — rare but valid if two
+                    // phones are on the same CGNAT-hairpinned
+                    // carrier LAN (some hotspot setups).
+                    out.push(SocketAddr::new(std::net::IpAddr::V4(v4), v4_port));
+                }
+            }
+            std::net::IpAddr::V6(v6) => {
+                // Phase 7: IPv6 host candidates via dedicated
+                // IPv6 socket. When v6_port is None, no IPv6
+                // endpoint exists — skip silently.
+                let Some(port) = v6_port else { continue };
+                if v6.is_loopback() || v6.is_unspecified() {
+                    continue;
+                }
+                // fe80::/10 link-local — needs scope ID, not
+                // routable across interfaces.
+                if (v6.segments()[0] & 0xffc0) == 0xfe80 {
+                    continue;
+                }
+                // Accept global unicast (2000::/3) and
+                // unique-local (fc00::/7).
+                let first_seg = v6.segments()[0];
+                let is_global = (first_seg & 0xe000) == 0x2000;
+                let is_ula = (first_seg & 0xfe00) == 0xfc00;
+                if is_global || is_ula {
+                    out.push(SocketAddr::new(std::net::IpAddr::V6(v6), port));
+                }
+            }
+        }
+    }
+    out
+}
+
+/// Role assignment for the Phase 3.5 dual-path QUIC race.
+///
+/// Both peers already know two strings at CallSetup time: their
+/// own server-reflexive address (queried via Phase 1 Reflect) and
+/// the peer's (carried in `CallSetup.peer_direct_addr`). To avoid
+/// a negotiation round-trip, both sides compare the two strings
+/// lexicographically and agree on a deterministic role:
+///
+/// - **Acceptor** — lexicographically smaller addr. Listens for
+///   an incoming direct connection from the peer. Does NOT dial.
+/// - **Dialer**   — lexicographically larger addr. Dials the
+///   peer's direct addr. Does NOT listen.
+///
+/// Both roles ALSO dial the relay in parallel as a fallback.
+/// Whichever future (direct or relay) completes first is used as
+/// the media transport. Because the role is deterministic and
+/// symmetric, both peers end up holding the same underlying QUIC
+/// session on the direct path — A's accepted conn and D's dialed
+/// conn are literally the same connection.
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub enum Role {
+    /// This peer listens for the direct incoming connection.
+    Acceptor,
+    /// This peer dials the peer's direct address.
+    Dialer,
+}
+
+/// Compute the deterministic role for this peer in the dual-path
+/// race. Returns `None` when no direct attempt is possible —
+/// either peer didn't advertise a reflex addr, or the two addrs
+/// are identical (same host on loopback / mis-advertised).
+///
+/// The caller should treat `None` as "skip direct, relay-only".
+pub fn determine_role(
+    own_reflex_addr: Option<&str>,
+    peer_reflex_addr: Option<&str>,
+) -> Option<Role> {
+    let (own, peer) = match (own_reflex_addr, peer_reflex_addr) {
+        (Some(o), Some(p)) => (o, p),
+        _ => return None,
+    };
+    match own.cmp(peer) {
+        std::cmp::Ordering::Less => Some(Role::Acceptor),
+        std::cmp::Ordering::Greater => Some(Role::Dialer),
+        // Equal addrs should never happen in production (both
+        // peers behind the same NAT mapping + same port would be
+        // a degenerate case). Guard against it so we don't infinite-
+        // loop waiting for a connection to ourselves.
+        std::cmp::Ordering::Equal => None,
+    }
+}
+
+/// Returns `true` if the address is in an RFC1918 / link-local /
+/// loopback range and therefore cannot possibly be a post-NAT
+/// reflex address from the public internet's point of view.
+///
+/// A probe against a relay ON THE SAME LAN as the client will
+/// naturally report the client's LAN IP back (because there's no
+/// NAT between them) — that observation is real but says nothing
+/// about the client's public-internet-facing NAT state. Mixing
+/// LAN reflex addrs with public-internet reflex addrs in
+/// `classify_nat` would always report `Multiple` (different IPs)
+/// and falsely warn about symmetric NAT. Filter them out before
+/// classifying.
+fn is_private_or_loopback(addr: &SocketAddr) -> bool {
+    match addr.ip() {
+        std::net::IpAddr::V4(v4) => {
+            let o = v4.octets();
+            v4.is_loopback()
+                || v4.is_private() // 10/8, 172.16/12, 192.168/16
+                || v4.is_link_local() // 169.254/16
+                || (o[0] == 100 && (o[1] & 0xc0) == 0x40) // 100.64/10 CGNAT shared
+        }
+        std::net::IpAddr::V6(v6) => {
+            v6.is_loopback() || v6.is_unspecified() || (v6.segments()[0] & 0xffc0) == 0xfe80 // fe80::/10 link-local
+        }
+    }
+}
+
+/// Pure-function NAT classifier — split out for unit testing
+/// without touching the network.
+///
+/// Only considers probes whose reflex addr is a **public-internet**
+/// address. LAN / private / loopback reflex addrs are dropped
+/// because they reflect the same-network path rather than the
+/// real NAT state. CGNAT (100.64/10) is also treated as private
+/// because the post-CGNAT address would be what we actually want
+/// to classify on — but CGNAT is unreachable from outside the
+/// carrier, so a relay seeing the CGNAT addr is on the same
+/// carrier network and again not useful for classification.
+pub fn classify_nat(probes: &[NatProbeResult]) -> (NatType, Option<String>) {
+    // First: parse every successful probe's observed addr.
+    let parsed: Vec<SocketAddr> = probes
+        .iter()
+        .filter_map(|p| p.observed_addr.as_deref().and_then(|s| s.parse().ok()))
+        .collect();
+
+    // Then: drop LAN / private / loopback reflex addrs. Those are
+    // legitimate observations by same-network relays, but they
+    // don't contribute to NAT-type classification because the
+    // client's real public-facing NAT mapping is not involved on
+    // that path. A relay on the same LAN always sees the client's
+    // LAN IP, regardless of whether the NAT beyond it is cone or
+    // symmetric.
+    let successes: Vec<SocketAddr> = parsed
+        .into_iter()
+        .filter(|a| !is_private_or_loopback(a))
+        .collect();
+
+    if successes.len() < 2 {
+        return (NatType::Unknown, None);
+    }
+
+    let first = successes[0];
+    let same_ip = successes.iter().all(|a| a.ip() == first.ip());
+    if !same_ip {
+        return (NatType::Multiple, None);
+    }
+
+    let same_port = successes.iter().all(|a| a.port() == first.port());
+    if same_port {
+        (NatType::Cone, Some(first.to_string()))
+    } else {
+        (NatType::SymmetricPort, None)
+    }
+}
+
+/// Enhanced NAT detection that combines relay-based reflection with
+/// public STUN server probes for more robust classification.
+///
+/// Runs both probe sets concurrently:
+/// 1. Relay probes via `detect_nat_type` (existing behavior)
+/// 2. Public STUN probes via `probe_stun_servers`
+///
+/// Merges all results and classifies. More probes = higher confidence
+/// in the NAT type classification. Falls back gracefully: if STUN
+/// servers are unreachable, relay probes still work (and vice versa).
+pub async fn detect_nat_type_with_stun(
+    relays: Vec<(String, SocketAddr)>,
+    timeout_ms: u64,
+    shared_endpoint: Option<wzp_transport::Endpoint>,
+    stun_config: &crate::stun::StunConfig,
+) -> NatDetection {
+    // Run relay probes and STUN probes concurrently.
+    let relay_fut = detect_nat_type(relays, timeout_ms, shared_endpoint);
+    let stun_fut = crate::stun::probe_stun_servers(stun_config);
+
+    let (relay_detection, stun_probes) = tokio::join!(relay_fut, stun_fut);
+
+    // Merge all probes and re-classify.
+    let mut all_probes = relay_detection.probes;
+    all_probes.extend(stun_probes);
+
+    let (nat_type, consensus_addr) = classify_nat(&all_probes);
+    NatDetection {
+        probes: all_probes,
+        nat_type,
+        consensus_addr,
+    }
+}
+
+// ── Unit tests for the pure classifier ───────────────────────────
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn mk(addr: Option<&str>) -> NatProbeResult {
+        NatProbeResult {
+            relay_name: "test".into(),
+            relay_addr: "0.0.0.0:0".into(),
+            observed_addr: addr.map(|s| s.to_string()),
+            latency_ms: addr.map(|_| 10),
+            error: None,
+        }
+    }
+
+    #[test]
+    fn classify_empty_is_unknown() {
+        let (nt, addr) = classify_nat(&[]);
+        assert_eq!(nt, NatType::Unknown);
+        assert!(addr.is_none());
+    }
+
+    #[test]
+    fn classify_single_success_is_unknown() {
+        let probes = vec![mk(Some("192.0.2.1:4433"))];
+        let (nt, addr) = classify_nat(&probes);
+        assert_eq!(nt, NatType::Unknown);
+        assert!(addr.is_none());
+    }
+
+    #[test]
+    fn classify_two_identical_is_cone() {
+        let probes = vec![
+            mk(Some("192.0.2.1:4433")),
+            mk(Some("192.0.2.1:4433")),
+        ];
+        let (nt, addr) = classify_nat(&probes);
+        assert_eq!(nt, NatType::Cone);
+        assert_eq!(addr.as_deref(), Some("192.0.2.1:4433"));
+    }
+
+    #[test]
+    fn classify_same_ip_different_ports_is_symmetric() {
+        let probes = vec![
+            mk(Some("192.0.2.1:4433")),
+            mk(Some("192.0.2.1:51234")),
+        ];
+        let (nt, addr) = classify_nat(&probes);
+        assert_eq!(nt, NatType::SymmetricPort);
+        assert!(addr.is_none());
+    }
+
+    #[test]
+    fn classify_different_ips_is_multiple() {
+        let probes = vec![
+            mk(Some("192.0.2.1:4433")),
+            mk(Some("198.51.100.9:4433")),
+        ];
+        let (nt, addr) = classify_nat(&probes);
+        assert_eq!(nt, NatType::Multiple);
+        assert!(addr.is_none());
+    }
+
+    #[test]
+    fn classify_drops_private_ip_probes() {
+        // One LAN probe + one public probe should behave like a
+        // single public probe — i.e. Unknown (not enough data to
+        // classify). This is the common real-world case: the user
+        // has a LAN relay + an internet relay configured, the LAN
+        // relay sees the LAN IP, the internet relay sees the WAN
+        // IP, and the old classifier would flag "Multiple" and
+        // falsely warn about symmetric NAT.
+        let probes = vec![
+            mk(Some("192.168.1.100:4433")), // LAN — must be dropped
+            mk(Some("203.0.113.5:4433")),   // public (TEST-NET-3)
+        ];
+        let (nt, _) = classify_nat(&probes);
+        assert_eq!(nt, NatType::Unknown);
+    }
+
+    #[test]
+    fn classify_drops_loopback_probes() {
+        let probes = vec![
+            mk(Some("127.0.0.1:4433")),     // loopback — must be dropped
+            mk(Some("203.0.113.5:4433")),   // public
+            mk(Some("203.0.113.5:4433")),   // public, same addr
+        ];
+        let (nt, addr) = classify_nat(&probes);
+        // Two public probes with identical addrs → Cone.
+        assert_eq!(nt, NatType::Cone);
+        assert_eq!(addr.as_deref(), Some("203.0.113.5:4433"));
+    }
+
+    #[test]
+    fn classify_drops_cgnat_probes() {
+        // 100.64.0.0/10 is the CGNAT shared-transition range.
+        // Filter treats it like RFC1918 — a relay that sees the
+        // client with a 100.64/10 addr is on the same CGNAT
+        // network and can't contribute to public NAT classification.
+        let probes = vec![
+            mk(Some("100.64.0.42:4433")),   // CGNAT — dropped
+            mk(Some("203.0.113.5:4433")),   // public
+            mk(Some("203.0.113.5:12345")),  // public, different port
+        ];
+        let (nt, _) = classify_nat(&probes);
+        // Two public probes same IP different port → SymmetricPort.
+        assert_eq!(nt, NatType::SymmetricPort);
+    }
+
+    #[test]
+    fn classify_two_lan_probes_is_unknown_not_cone() {
+        // Even if both probes come back from LAN relays, we can't
+        // say anything useful about the public NAT state. Unknown,
+        // not Cone.
+        let probes = vec![
+            mk(Some("192.168.1.100:4433")),
+            mk(Some("192.168.1.100:4433")),
+        ];
+        let (nt, addr) = classify_nat(&probes);
+        assert_eq!(nt, NatType::Unknown);
+        assert!(addr.is_none());
+    }
+
+    #[test]
+    fn classify_mix_of_success_and_failure() {
+        let probes = vec![
+            mk(Some("192.0.2.1:4433")),
+            mk(None), // failed probe
+            mk(Some("192.0.2.1:4433")),
+        ];
+        let (nt, addr) = classify_nat(&probes);
+        // Two successes both agree → Cone, ignore the failure row.
+        assert_eq!(nt, NatType::Cone);
+        assert_eq!(addr.as_deref(), Some("192.0.2.1:4433"));
+    }
+
+    #[test]
+    fn determine_role_smaller_is_acceptor() {
+        // Lexicographic: "192.0.2.1:4433" < "198.51.100.9:4433"
+        assert_eq!(
+            determine_role(Some("192.0.2.1:4433"), Some("198.51.100.9:4433")),
+            Some(Role::Acceptor)
+        );
+    }
+
+    #[test]
+    fn determine_role_larger_is_dialer() {
+        assert_eq!(
+            determine_role(Some("198.51.100.9:4433"), Some("192.0.2.1:4433")),
+            Some(Role::Dialer)
+        );
+    }
+
+    #[test]
+    fn determine_role_port_difference_matters() {
+        // Same ip, different ports — string compare still works
+        // because "4433" < "54321".
+        assert_eq!(
+            determine_role(Some("127.0.0.1:4433"), Some("127.0.0.1:54321")),
+            Some(Role::Acceptor)
+        );
+        assert_eq!(
+            determine_role(Some("127.0.0.1:54321"), Some("127.0.0.1:4433")),
+            Some(Role::Dialer)
+        );
+    }
+
+    #[test]
+    fn determine_role_equal_addrs_is_none() {
+        assert_eq!(
+            determine_role(Some("192.0.2.1:4433"), Some("192.0.2.1:4433")),
+            None
+        );
+    }
+
+    #[test]
+    fn determine_role_missing_side_is_none() {
+        assert_eq!(determine_role(None, Some("192.0.2.1:4433")), None);
+        assert_eq!(determine_role(Some("192.0.2.1:4433"), None), None);
+        assert_eq!(determine_role(None, None), None);
+    }
+
+    #[test]
+    fn determine_role_is_symmetric_across_peers() {
+        // Both peers compute roles independently; they must end
+        // up with opposite assignments (one Acceptor, one Dialer)
+        // so that each side ends up talking to the other.
+        let a = "192.0.2.1:4433";
+        let b = "198.51.100.9:4433";
+        let alice_role = determine_role(Some(a), Some(b));
+        let bob_role = determine_role(Some(b), Some(a));
+        assert_eq!(alice_role, Some(Role::Acceptor));
+        assert_eq!(bob_role, Some(Role::Dialer));
+    }
+
+    #[test]
+    fn classify_one_success_one_failure_is_unknown() {
+        let probes = vec![mk(Some("192.0.2.1:4433")), mk(None)];
+        let (nt, addr) = classify_nat(&probes);
+        assert_eq!(nt, NatType::Unknown);
+        assert!(addr.is_none());
+    }
+}

crates/wzp-client/src/relay_map.rs

@@ -0,0 +1,339 @@
+//! Phase 8 (Tailscale-inspired): Relay map for automatic relay
+//! selection based on latency.
+//!
+//! Maintains a sorted list of known relays with their measured
+//! latencies. Used during call setup to pick the lowest-latency
+//! relay, and by netcheck to report relay health.
+
+use std::net::SocketAddr;
+use std::time::{Duration, Instant};
+
+use serde::Serialize;
+
+/// A known relay endpoint with measured latency.
+#[derive(Debug, Clone, Serialize)]
+pub struct RelayEntry {
+    /// Human-readable name (e.g., "us-east", "eu-west").
+    pub name: String,
+    /// Relay address.
+    pub addr: SocketAddr,
+    /// Geographic region (from RegisterPresenceAck).
+    pub region: Option<String>,
+    /// Last measured RTT (ms).
+    pub rtt_ms: Option<u32>,
+    /// When the RTT was last measured.
+    #[serde(skip)]
+    pub last_probed: Option<Instant>,
+    /// Whether this relay is currently reachable.
+    pub reachable: bool,
+}
+
+/// Sorted relay map. Entries are ordered by RTT (lowest first).
+#[derive(Debug, Clone, Default)]
+pub struct RelayMap {
+    entries: Vec<RelayEntry>,
+}
+
+impl RelayMap {
+    pub fn new() -> Self {
+        Self {
+            entries: Vec::new(),
+        }
+    }
+
+    /// Add or update a relay entry.
+    pub fn upsert(&mut self, name: &str, addr: SocketAddr, region: Option<String>) {
+        if let Some(entry) = self.entries.iter_mut().find(|e| e.addr == addr) {
+            entry.name = name.to_string();
+            if region.is_some() {
+                entry.region = region;
+            }
+        } else {
+            self.entries.push(RelayEntry {
+                name: name.to_string(),
+                addr,
+                region,
+                rtt_ms: None,
+                last_probed: None,
+                reachable: false,
+            });
+        }
+    }
+
+    /// Update RTT measurement for a relay.
+    pub fn update_rtt(&mut self, addr: SocketAddr, rtt_ms: u32) {
+        if let Some(entry) = self.entries.iter_mut().find(|e| e.addr == addr) {
+            entry.rtt_ms = Some(rtt_ms);
+            entry.last_probed = Some(Instant::now());
+            entry.reachable = true;
+        }
+        self.sort();
+    }
+
+    /// Mark a relay as unreachable.
+    pub fn mark_unreachable(&mut self, addr: SocketAddr) {
+        if let Some(entry) = self.entries.iter_mut().find(|e| e.addr == addr) {
+            entry.reachable = false;
+            entry.last_probed = Some(Instant::now());
+        }
+        self.sort();
+    }
+
+    /// Get the preferred (lowest-latency, reachable) relay.
+    pub fn preferred(&self) -> Option<&RelayEntry> {
+        self.entries
+            .iter()
+            .find(|e| e.reachable && e.rtt_ms.is_some())
+    }
+
+    /// Get all entries, sorted by RTT.
+    pub fn entries(&self) -> &[RelayEntry] {
+        &self.entries
+    }
+
+    /// Populate from a `RegisterPresenceAck.available_relays` list.
+    /// Each entry is "name|addr" format.
+    pub fn populate_from_ack(&mut self, relays: &[String], relay_region: Option<&str>) {
+        for entry_str in relays {
+            if let Some((name, addr_str)) = entry_str.split_once('|') {
+                if let Ok(addr) = addr_str.parse::<SocketAddr>() {
+                    self.upsert(name, addr, None);
+                }
+            }
+        }
+        // If the ack included a region for the current relay, we
+        // could tag it — but we'd need to know which relay we're
+        // connected to. Left for the caller to handle.
+        let _ = relay_region;
+    }
+
+    /// Check if any entry has a stale probe (older than `max_age`).
+    pub fn needs_reprobe(&self, max_age: Duration) -> bool {
+        self.entries.iter().any(|e| {
+            match e.last_probed {
+                None => true,
+                Some(t) => t.elapsed() > max_age,
+            }
+        })
+    }
+
+    /// Get entries that need reprobing.
+    pub fn stale_entries(&self, max_age: Duration) -> Vec<(String, SocketAddr)> {
+        self.entries
+            .iter()
+            .filter(|e| match e.last_probed {
+                None => true,
+                Some(t) => t.elapsed() > max_age,
+            })
+            .map(|e| (e.name.clone(), e.addr))
+            .collect()
+    }
+
+    fn sort(&mut self) {
+        self.entries.sort_by_key(|e| {
+            if e.reachable {
+                e.rtt_ms.unwrap_or(u32::MAX)
+            } else {
+                u32::MAX
+            }
+        });
+    }
+}
+
+// ── Tests ──────────────────────────────────────────────────────────
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn preferred_returns_lowest_rtt() {
+        let mut map = RelayMap::new();
+        let a1: SocketAddr = "10.0.0.1:4433".parse().unwrap();
+        let a2: SocketAddr = "10.0.0.2:4433".parse().unwrap();
+        let a3: SocketAddr = "10.0.0.3:4433".parse().unwrap();
+
+        map.upsert("slow", a1, None);
+        map.upsert("fast", a2, None);
+        map.upsert("mid", a3, None);
+
+        map.update_rtt(a1, 200);
+        map.update_rtt(a2, 15);
+        map.update_rtt(a3, 80);
+
+        let pref = map.preferred().unwrap();
+        assert_eq!(pref.addr, a2);
+        assert_eq!(pref.rtt_ms, Some(15));
+    }
+
+    #[test]
+    fn unreachable_not_preferred() {
+        let mut map = RelayMap::new();
+        let a1: SocketAddr = "10.0.0.1:4433".parse().unwrap();
+        let a2: SocketAddr = "10.0.0.2:4433".parse().unwrap();
+
+        map.upsert("fast-dead", a1, None);
+        map.upsert("slow-alive", a2, None);
+
+        map.update_rtt(a1, 5);
+        map.update_rtt(a2, 200);
+        map.mark_unreachable(a1);
+
+        let pref = map.preferred().unwrap();
+        assert_eq!(pref.addr, a2);
+    }
+
+    #[test]
+    fn populate_from_ack() {
+        let mut map = RelayMap::new();
+        map.populate_from_ack(
+            &[
+                "us-east|203.0.113.5:4433".into(),
+                "eu-west|198.51.100.9:4433".into(),
+            ],
+            Some("us-east"),
+        );
+        assert_eq!(map.entries().len(), 2);
+        assert_eq!(map.entries()[0].name, "us-east");
+        assert_eq!(map.entries()[1].name, "eu-west");
+    }
+
+    #[test]
+    fn upsert_updates_existing() {
+        let mut map = RelayMap::new();
+        let addr: SocketAddr = "10.0.0.1:4433".parse().unwrap();
+        map.upsert("old-name", addr, None);
+        map.upsert("new-name", addr, Some("us-west".into()));
+        assert_eq!(map.entries().len(), 1);
+        assert_eq!(map.entries()[0].name, "new-name");
+        assert_eq!(map.entries()[0].region, Some("us-west".into()));
+    }
+
+    #[test]
+    fn upsert_preserves_region_when_none() {
+        let mut map = RelayMap::new();
+        let addr: SocketAddr = "10.0.0.1:4433".parse().unwrap();
+        map.upsert("relay", addr, Some("eu-west".into()));
+        map.upsert("relay", addr, None); // region is None
+        // Should keep the original region
+        assert_eq!(map.entries()[0].region, Some("eu-west".into()));
+    }
+
+    #[test]
+    fn preferred_returns_none_on_empty() {
+        let map = RelayMap::new();
+        assert!(map.preferred().is_none());
+    }
+
+    #[test]
+    fn preferred_returns_none_when_all_unreachable() {
+        let mut map = RelayMap::new();
+        let addr: SocketAddr = "10.0.0.1:4433".parse().unwrap();
+        map.upsert("relay", addr, None);
+        // Not update_rtt'd, so reachable=false
+        assert!(map.preferred().is_none());
+    }
+
+    #[test]
+    fn needs_reprobe_empty_is_false() {
+        let map = RelayMap::new();
+        // No entries → nothing to reprobe
+        assert!(!map.needs_reprobe(Duration::from_secs(60)));
+    }
+
+    #[test]
+    fn needs_reprobe_never_probed() {
+        let mut map = RelayMap::new();
+        map.upsert("relay", "10.0.0.1:4433".parse().unwrap(), None);
+        assert!(map.needs_reprobe(Duration::from_secs(60)));
+    }
+
+    #[test]
+    fn needs_reprobe_fresh_is_false() {
+        let mut map = RelayMap::new();
+        let addr: SocketAddr = "10.0.0.1:4433".parse().unwrap();
+        map.upsert("relay", addr, None);
+        map.update_rtt(addr, 50);
+        // Just probed, so 60s max_age should not trigger
+        assert!(!map.needs_reprobe(Duration::from_secs(60)));
+    }
+
+    #[test]
+    fn stale_entries_returns_unprobed() {
+        let mut map = RelayMap::new();
+        let a1: SocketAddr = "10.0.0.1:4433".parse().unwrap();
+        let a2: SocketAddr = "10.0.0.2:4433".parse().unwrap();
+        map.upsert("probed", a1, None);
+        map.upsert("stale", a2, None);
+        map.update_rtt(a1, 50);
+
+        let stale = map.stale_entries(Duration::from_secs(60));
+        assert_eq!(stale.len(), 1);
+        assert_eq!(stale[0].1, a2);
+    }
+
+    #[test]
+    fn sort_stability_with_equal_rtt() {
+        let mut map = RelayMap::new();
+        let a1: SocketAddr = "10.0.0.1:4433".parse().unwrap();
+        let a2: SocketAddr = "10.0.0.2:4433".parse().unwrap();
+        map.upsert("first", a1, None);
+        map.upsert("second", a2, None);
+        map.update_rtt(a1, 50);
+        map.update_rtt(a2, 50);
+
+        // Both have same RTT — sort should be stable (insertion order)
+        assert_eq!(map.entries().len(), 2);
+        // Both are valid preferred relays
+        assert!(map.preferred().is_some());
+    }
+
+    #[test]
+    fn populate_from_ack_skips_malformed() {
+        let mut map = RelayMap::new();
+        map.populate_from_ack(
+            &[
+                "good|10.0.0.1:4433".into(),
+                "no-pipe-separator".into(),
+                "bad-addr|not-a-socket-addr".into(),
+                "also-good|10.0.0.2:4433".into(),
+            ],
+            None,
+        );
+        assert_eq!(map.entries().len(), 2);
+    }
+
+    #[test]
+    fn mark_unreachable_sorts_to_end() {
+        let mut map = RelayMap::new();
+        let a1: SocketAddr = "10.0.0.1:4433".parse().unwrap();
+        let a2: SocketAddr = "10.0.0.2:4433".parse().unwrap();
+        map.upsert("fast", a1, None);
+        map.upsert("slow", a2, None);
+        map.update_rtt(a1, 10);
+        map.update_rtt(a2, 200);
+
+        assert_eq!(map.preferred().unwrap().addr, a1);
+
+        map.mark_unreachable(a1);
+        assert_eq!(map.preferred().unwrap().addr, a2);
+    }
+
+    #[test]
+    fn relay_entry_serializes() {
+        let entry = RelayEntry {
+            name: "test".into(),
+            addr: "10.0.0.1:4433".parse().unwrap(),
+            region: Some("us-east".into()),
+            rtt_ms: Some(42),
+            last_probed: Some(Instant::now()),
+            reachable: true,
+        };
+        let json = serde_json::to_string(&entry).unwrap();
+        assert!(json.contains("test"));
+        assert!(json.contains("us-east"));
+        assert!(json.contains("42"));
+        // last_probed is #[serde(skip)]
+        assert!(!json.contains("last_probed"));
+    }
+}

crates/wzp-client/tests/dual_path.rs

@@ -0,0 +1,222 @@
+//! Phase 3.5 integration tests for the dual-path QUIC race.
+//!
+//! The race takes a role (Acceptor or Dialer), a peer_direct_addr,
+//! a relay_addr, and two SNI strings, then returns whichever QUIC
+//! handshake completes first wrapped in a `QuinnTransport`. These
+//! tests validate that:
+//!
+//! 1. On loopback with two real clients playing A + D roles, the
+//!    direct path wins (fewer hops than relay).
+//! 2. When the direct peer is dead (nothing listening) but the
+//!    relay is up, the relay wins within the fallback window.
+//! 3. When both paths are dead, the race errors cleanly rather
+//!    than hanging forever.
+//!
+//! The "relay" in these tests is a minimal mock that just accepts
+//! an incoming QUIC connection and drops it — we don't need any
+//! protocol handling, just a TCP-ish listen-and-accept.
+
+use std::net::{Ipv4Addr, SocketAddr};
+use std::time::Duration;
+
+use wzp_client::dual_path::{race, PeerCandidates, WinningPath};
+use wzp_client::reflect::Role;
+use wzp_transport::{create_endpoint, server_config};
+
+/// Spin up a "relay-ish" mock server on loopback that accepts
+/// incoming QUIC connections and does nothing with them. Used to
+/// give the relay branch of the race a real target to dial.
+/// Returns the bound address + a join handle (kept alive to keep
+/// the endpoint up).
+async fn spawn_mock_relay() -> (SocketAddr, tokio::task::JoinHandle<()>) {
+    let _ = rustls::crypto::ring::default_provider().install_default();
+    let (sc, _cert_der) = server_config();
+    let bind: SocketAddr = (Ipv4Addr::LOCALHOST, 0).into();
+    let ep = create_endpoint(bind, Some(sc)).expect("relay endpoint");
+    let addr = ep.local_addr().expect("local_addr");
+
+    let handle = tokio::spawn(async move {
+        // Accept loop — hold the connection alive for a short
+        // while so the race result isn't killed by the peer
+        // closing before the winning transport is returned.
+        while let Some(incoming) = ep.accept().await {
+            if let Ok(_conn) = incoming.await {
+                tokio::time::sleep(Duration::from_secs(5)).await;
+            }
+        }
+    });
+    (addr, handle)
+}
+
+// -----------------------------------------------------------------------
+// Test 1: direct path wins when both sides are up
+// -----------------------------------------------------------------------
+//
+// Spawn a mock relay, then set up a two-client test where one
+// client plays the Acceptor role and the other plays the Dialer
+// role. The Dialer's `peer_direct_addr` is the Acceptor's listen
+// address. Because the direct path is a single loopback hop and
+// the relay dial also terminates on loopback, both complete
+// essentially instantly — the `biased` tokio::select in race()
+// should pick direct.
+
+#[tokio::test(flavor = "multi_thread", worker_threads = 4)]
+async fn dual_path_direct_wins_on_loopback() {
+    let _ = rustls::crypto::ring::default_provider().install_default();
+    let (relay_addr, _relay_handle) = spawn_mock_relay().await;
+
+    // Acceptor task: run race(Role::Acceptor, peer_addr_placeholder, ...).
+    // Since the acceptor doesn't dial, the peer_direct_addr arg is
+    // unused on the direct branch but we still pass a placeholder
+    // because the API takes one. Use a stub addr that would error
+    // if it were ever dialed — proving the Acceptor really doesn't
+    // reach it.
+    let unused_addr: SocketAddr = "127.0.0.1:2".parse().unwrap();
+
+    // We can't race both sides in the same task because each race
+    // call has its own direct endpoint that needs to talk to the
+    // OTHER side's endpoint. So spawn the Acceptor in a task and
+    // let it expose its listen addr via a oneshot back to the test,
+    // then run the Dialer in the test's main task.
+    //
+    // There's a chicken-and-egg issue: the Acceptor's listen addr
+    // is only known after race() creates its endpoint. To avoid
+    // reaching into race()'s internals, we instead play a slight
+    // trick: create the Acceptor's endpoint ourselves (outside
+    // race()) to learn its addr, spin up an accept loop on it
+    // ourselves, and pass THAT addr as the Dialer's peer addr.
+    // This tests the Dialer->Acceptor handshake end-to-end without
+    // running the full race() on both sides.
+
+    let (sc, _cert_der) = server_config();
+    let acceptor_bind: SocketAddr = (Ipv4Addr::LOCALHOST, 0).into();
+    let acceptor_ep = create_endpoint(acceptor_bind, Some(sc)).expect("acceptor ep");
+    let acceptor_listen_addr = acceptor_ep.local_addr().expect("acceptor addr");
+
+    // Drop the external acceptor after the test finishes, not
+    // before — spawn a dedicated accept task.
+    let acceptor_accept_task = tokio::spawn(async move {
+        // Accept one connection and hold it for a while so the
+        // Dialer side can complete its QUIC handshake.
+        if let Some(incoming) = acceptor_ep.accept().await {
+            if let Ok(_conn) = incoming.await {
+                tokio::time::sleep(Duration::from_secs(5)).await;
+            }
+        }
+    });
+
+    // Now run the Dialer in the race — peer_direct_addr = acceptor's
+    // listen addr. The relay is the mock from above. Direct path
+    // should win.
+    let result = race(
+        Role::Dialer,
+        PeerCandidates {
+            reflexive: Some(acceptor_listen_addr),
+            local: Vec::new(),
+            mapped: None,
+        },
+        relay_addr,
+        "test-room".into(),
+        "call-test".into(),
+        None, // own_reflexive: not needed in tests
+        None, // Phase 5: tests use fresh endpoints (no shared signal)
+        None, // Phase 7: no IPv6 endpoint in tests
+    )
+    .await
+    .expect("race must succeed");
+
+    assert!(result.direct_transport.is_some(), "direct transport should be available");
+    assert_eq!(result.local_winner, WinningPath::Direct, "direct should win on loopback");
+
+    // Cancel the acceptor accept task so the test finishes.
+    acceptor_accept_task.abort();
+    // Suppress unused-var warning for the placeholder.
+    let _ = unused_addr;
+}
+
+// -----------------------------------------------------------------------
+// Test 2: relay wins when the direct peer is dead
+// -----------------------------------------------------------------------
+//
+// Dialer role, peer_direct_addr = a port nothing is listening on,
+// relay is the working mock. Direct dial will sit waiting for a
+// QUIC handshake that never comes; the 2s direct timeout kicks in
+// and the relay path wins the fallback.
+
+#[tokio::test(flavor = "multi_thread", worker_threads = 4)]
+async fn dual_path_relay_wins_when_direct_is_dead() {
+    let _ = rustls::crypto::ring::default_provider().install_default();
+    let (relay_addr, _relay_handle) = spawn_mock_relay().await;
+
+    // A port that nothing is listening on — dead direct target.
+    // Port 1 on loopback is almost never bound and UDP packets to
+    // it will be dropped silently, so the QUIC handshake times out.
+    let dead_peer: SocketAddr = "127.0.0.1:1".parse().unwrap();
+
+    let result = race(
+        Role::Dialer,
+        PeerCandidates {
+            reflexive: Some(dead_peer),
+            local: Vec::new(),
+            mapped: None,
+        },
+        relay_addr,
+        "test-room".into(),
+        "call-test".into(),
+        None, // own_reflexive: not needed in tests
+        None, // Phase 5: tests use fresh endpoints (no shared signal)
+        None, // Phase 7: no IPv6 endpoint in tests
+    )
+    .await
+    .expect("race must succeed via relay fallback");
+
+    assert!(result.relay_transport.is_some(), "relay transport should be available");
+    assert_eq!(
+        result.local_winner,
+        WinningPath::Relay,
+        "relay should win when direct dial has nowhere to land"
+    );
+}
+
+// -----------------------------------------------------------------------
+// Test 3: race errors cleanly when both paths are dead
+// -----------------------------------------------------------------------
+//
+// Dialer role, peer_direct_addr = dead, relay_addr = dead.
+// Expected: race returns an Err within ~7s (2s direct timeout +
+// 5s relay timeout fallback).
+
+#[tokio::test(flavor = "multi_thread", worker_threads = 4)]
+async fn dual_path_errors_cleanly_when_both_paths_dead() {
+    let _ = rustls::crypto::ring::default_provider().install_default();
+
+    let dead_peer: SocketAddr = "127.0.0.1:1".parse().unwrap();
+    let dead_relay: SocketAddr = "127.0.0.1:2".parse().unwrap();
+
+    let start = std::time::Instant::now();
+    let result = race(
+        Role::Dialer,
+        PeerCandidates {
+            reflexive: Some(dead_peer),
+            local: Vec::new(),
+            mapped: None,
+        },
+        dead_relay,
+        "test-room".into(),
+        "call-test".into(),
+        None, // own_reflexive: not needed in tests
+        None, // Phase 5: tests use fresh endpoints (no shared signal)
+        None, // Phase 7: no IPv6 endpoint in tests
+    )
+    .await;
+    let elapsed = start.elapsed();
+
+    assert!(result.is_err(), "both-dead must return Err");
+    // Upper bound: direct 2s timeout + relay 5s fallback + small
+    // slack for scheduling. If this blows, something is looping.
+    assert!(
+        elapsed < Duration::from_secs(10),
+        "race took too long to give up: {:?}",
+        elapsed
+    );
+}

crates/wzp-client/tests/handshake_integration.rs

@@ -83,12 +83,12 @@ async fn full_handshake_both_sides_derive_same_session() {
 
     // Run client and relay handshakes concurrently.
     let (client_result, relay_result) = tokio::join!(
-        wzp_client::handshake::perform_handshake(client_transport_clone.as_ref(), &client_seed),
+        wzp_client::handshake::perform_handshake(client_transport_clone.as_ref(), &client_seed, None),
         wzp_relay::handshake::accept_handshake(relay_transport_clone.as_ref(), &relay_seed),
     );
 
     let mut client_session = client_result.expect("client handshake should succeed");
-    let (mut relay_session, chosen_profile) =
+    let (mut relay_session, chosen_profile, _caller_fp, _caller_alias) =
         relay_result.expect("relay handshake should succeed");
 
     // Verify a profile was chosen.
@@ -151,6 +151,7 @@ async fn handshake_rejects_tampered_signature() {
             ephemeral_pub,
             signature: bad_signature,
             supported_profiles: vec![wzp_proto::QualityProfile::GOOD],
+            alias: None,
         };
         client_transport_clone
             .send_signal(&offer)

crates/wzp-codec/Cargo.toml

@@ -10,8 +10,17 @@ description = "WarzonePhone audio codec layer — Opus + Codec2 encoding/decodin
 wzp-proto = { workspace = true }
 tracing = { workspace = true }
 
-# Opus bindings
-audiopus = { workspace = true }
+# Opus bindings — libopus 1.5.2.
+# opusic-c for the encoder (set_dred_duration lives here in Phase 1).
+# opusic-sys for the decoder — we wrap the raw *mut OpusDecoder ourselves
+# because opusic-c::Decoder.inner is pub(crate), blocking the unified
+# decoder + DRED path we need in Phase 3.
+opusic-c = { workspace = true }
+opusic-sys = { workspace = true }
+
+# Zero-cost slice reinterpretation for the i16 ↔ u16 boundary between
+# our PCM buffers and opusic-c's encode API.
+bytemuck = { workspace = true }
 
 # Pure-Rust Codec2 implementation
 codec2 = { workspace = true }

crates/wzp-codec/src/adaptive.rs

@@ -116,6 +116,14 @@ impl AudioEncoder for AdaptiveEncoder {
     fn set_dtx(&mut self, enabled: bool) {
         self.opus.set_dtx(enabled);
     }
+
+    fn set_expected_loss(&mut self, loss_pct: u8) {
+        self.opus.set_expected_loss(loss_pct);
+    }
+
+    fn set_dred_duration(&mut self, frames: u8) {
+        self.opus.set_dred_duration(frames);
+    }
 }
 
 // ─── AdaptiveDecoder ─────────────────────────────────────────────────────────
@@ -199,6 +207,27 @@ impl AdaptiveDecoder {
     fn codec2_frame_samples(&self) -> usize {
         self.codec2.frame_samples()
     }
+
+    /// Reconstruct a lost frame from a previously parsed DRED state.
+    ///
+    /// Phase 3b entry point for gap reconstruction. Dispatches to the
+    /// inner Opus decoder when active. Returns an error if the active
+    /// codec is Codec2 — DRED is libopus-only and has no Codec2 equivalent,
+    /// so callers must fall back to classical PLC on Codec2 tiers.
+    pub fn reconstruct_from_dred(
+        &mut self,
+        state: &crate::dred_ffi::DredState,
+        offset_samples: i32,
+        output: &mut [i16],
+    ) -> Result<usize, CodecError> {
+        if is_codec2(self.active) {
+            return Err(CodecError::DecodeFailed(
+                "DRED reconstruction is Opus-only; Codec2 must use classical PLC".into(),
+            ));
+        }
+        self.opus
+            .reconstruct_from_dred(state, offset_samples, output)
+    }
 }
 
 // ─── Tests ───────────────────────────────────────────────────────────────────

crates/wzp-codec/src/aec.rs

@@ -1,53 +1,127 @@
-//! Acoustic Echo Cancellation using NLMS adaptive filter.
-//! Processes 480-sample (10ms) sub-frames at 48kHz.
+//! Acoustic Echo Cancellation — delay-compensated leaky NLMS with
+//! Geigel double-talk detection.
+//!
+//! Key insight: on a laptop, the round-trip audio latency (playout → speaker
+//! → air → mic → capture) is 30–50ms.  The far-end reference must be delayed
+//! by this amount so the adaptive filter models the *echo path*, not the
+//! *system delay + echo path*.
+//!
+//! The leaky coefficient decay prevents the filter from diverging when the
+//! echo path changes (e.g. hand near laptop) or when the delay estimate
+//! is slightly off.
 
-/// NLMS (Normalized Least Mean Squares) adaptive filter echo canceller.
-///
-/// Removes acoustic echo by modelling the echo path between the far-end
-/// (speaker) signal and the near-end (microphone) signal, then subtracting
-/// the estimated echo from the near-end in real time.
+/// Delay-compensated leaky NLMS echo canceller with Geigel DTD.
 pub struct EchoCanceller {
-    filter_coeffs: Vec<f32>,
+    // --- Adaptive filter ---
+    filter: Vec<f32>,
     filter_len: usize,
-    far_end_buf: Vec<f32>,
-    far_end_pos: usize,
+    /// Circular buffer of far-end reference samples (after delay).
+    far_buf: Vec<f32>,
+    far_pos: usize,
+    /// NLMS step size.
     mu: f32,
+    /// Leakage factor: coefficients are multiplied by (1 - leak) each frame.
+    /// Prevents unbounded growth / divergence.  0.0001 is gentle.
+    leak: f32,
     enabled: bool,
+
+    // --- Delay buffer ---
+    /// Raw far-end samples before delay compensation.
+    delay_ring: Vec<f32>,
+    delay_write: usize,
+    delay_read: usize,
+    /// Delay in samples (e.g. 1920 = 40ms at 48kHz).
+    delay_samples: usize,
+    /// Capacity of the delay ring.
+    delay_cap: usize,
+
+    // --- Double-talk detection (Geigel) ---
+    /// Peak far-end level over the last filter_len samples.
+    far_peak: f32,
+    /// Geigel threshold: if |near| > threshold * far_peak, assume double-talk.
+    geigel_threshold: f32,
+    /// Holdover counter: keep DTD active for a few frames after detection.
+    dtd_holdover: u32,
+    dtd_hold_frames: u32,
 }
 
 impl EchoCanceller {
     /// Create a new echo canceller.
     ///
     /// * `sample_rate` — typically 48000
-    /// * `filter_ms`   — echo-tail length in milliseconds (e.g. 100 for 100 ms)
+    /// * `filter_ms`   — echo-tail length in milliseconds (60ms recommended)
+    /// * `delay_ms`    — far-end delay compensation in milliseconds (40ms for laptops)
     pub fn new(sample_rate: u32, filter_ms: u32) -> Self {
+        Self::with_delay(sample_rate, filter_ms, 40)
+    }
+
+    pub fn with_delay(sample_rate: u32, filter_ms: u32, delay_ms: u32) -> Self {
         let filter_len = (sample_rate as usize) * (filter_ms as usize) / 1000;
+        let delay_samples = (sample_rate as usize) * (delay_ms as usize) / 1000;
+        // Delay ring must hold at least delay_samples + one frame (960) of headroom.
+        let delay_cap = delay_samples + (sample_rate as usize / 10); // +100ms headroom
         Self {
-            filter_coeffs: vec![0.0f32; filter_len],
+            filter: vec![0.0; filter_len],
             filter_len,
-            far_end_buf: vec![0.0f32; filter_len],
-            far_end_pos: 0,
+            far_buf: vec![0.0; filter_len],
+            far_pos: 0,
             mu: 0.01,
+            leak: 0.0001,
             enabled: true,
+
+            delay_ring: vec![0.0; delay_cap],
+            delay_write: 0,
+            delay_read: 0,
+            delay_samples,
+            delay_cap,
+
+            far_peak: 0.0,
+            geigel_threshold: 0.7,
+            dtd_holdover: 0,
+            dtd_hold_frames: 5,
         }
     }
 
-    /// Feed far-end (speaker/playback) samples into the circular buffer.
-    ///
-    /// Must be called with the audio that was played out through the speaker
-    /// *before* the corresponding near-end frame is processed.
+    /// Feed far-end (speaker) samples.  These go into the delay buffer first;
+    /// once enough samples have accumulated, they are released to the filter's
+    /// circular buffer with the correct delay offset.
     pub fn feed_farend(&mut self, farend: &[i16]) {
+        // Write raw samples into the delay ring.
         for &s in farend {
-            self.far_end_buf[self.far_end_pos] = s as f32;
-            self.far_end_pos = (self.far_end_pos + 1) % self.filter_len;
+            self.delay_ring[self.delay_write % self.delay_cap] = s as f32;
+            self.delay_write += 1;
+        }
+
+        // Release delayed samples to the filter's far-end buffer.
+        while self.delay_available() >= 1 {
+            let sample = self.delay_ring[self.delay_read % self.delay_cap];
+            self.delay_read += 1;
+
+            self.far_buf[self.far_pos] = sample;
+            self.far_pos = (self.far_pos + 1) % self.filter_len;
+
+            // Track peak far-end level for Geigel DTD.
+            let abs_s = sample.abs();
+            if abs_s > self.far_peak {
+                self.far_peak = abs_s;
+            }
+        }
+
+        // Decay far_peak slowly (avoids stale peak from a loud burst long ago).
+        self.far_peak *= 0.9995;
+    }
+
+    /// Number of delayed samples available to release.
+    fn delay_available(&self) -> usize {
+        let buffered = self.delay_write - self.delay_read;
+        if buffered > self.delay_samples {
+            buffered - self.delay_samples
+        } else {
+            0
         }
     }
 
     /// Process a near-end (microphone) frame, removing the estimated echo.
-    ///
-    /// Returns the echo-return-loss enhancement (ERLE) as a ratio: the RMS of
-    /// the original near-end divided by the RMS of the residual.  Values > 1.0
-    /// mean echo was reduced.
     pub fn process_frame(&mut self, nearend: &mut [i16]) -> f32 {
         if !self.enabled {
             return 1.0;
@@ -56,85 +130,96 @@ impl EchoCanceller {
         let n = nearend.len();
         let fl = self.filter_len;
 
+        // --- Geigel double-talk detection ---
+        // If any near-end sample exceeds threshold * far_peak, assume
+        // the local speaker is active and freeze adaptation.
+        let mut is_doubletalk = self.dtd_holdover > 0;
+        if !is_doubletalk {
+            let threshold_level = self.geigel_threshold * self.far_peak;
+            for &s in nearend.iter() {
+                if (s as f32).abs() > threshold_level && self.far_peak > 100.0 {
+                    is_doubletalk = true;
+                    self.dtd_holdover = self.dtd_hold_frames;
+                    break;
+                }
+            }
+        }
+        if self.dtd_holdover > 0 {
+            self.dtd_holdover -= 1;
+        }
+
+        // Check if far-end is active (otherwise nothing to cancel).
+        let far_active = self.far_peak > 100.0;
+
+        // --- Leaky coefficient decay ---
+        // Applied once per frame for efficiency.
+        let decay = 1.0 - self.leak;
+        for c in self.filter.iter_mut() {
+            *c *= decay;
+        }
+
         let mut sum_near_sq: f64 = 0.0;
         let mut sum_err_sq: f64 = 0.0;
 
         for i in 0..n {
             let near_f = nearend[i] as f32;
 
-            // --- estimate echo as dot(coeffs, farend_window) ---
-            // The far-end window for this sample starts at
-            //   (far_end_pos - 1 - i) mod filter_len   (most recent)
-            // and goes back filter_len samples.
+            // Position of far-end "now" for this near-end sample.
+            let base = (self.far_pos + fl * ((n / fl) + 2) + i - n) % fl;
+
+            // --- Echo estimation: dot(filter, far_end_window) ---
             let mut echo_est: f32 = 0.0;
             let mut power: f32 = 0.0;
 
-            // Position of the most-recent far-end sample for this near-end sample.
-            // far_end_pos points to the *next write* position, so the most-recent
-            // sample written is at far_end_pos - 1.  We have already called
-            // feed_farend for this block, so the relevant samples are the last
-            // filter_len entries ending just before the current write position,
-            // offset by how far we are into this near-end frame.
-            //
-            // For sample i of the near-end frame, the corresponding far-end
-            // "now" is far_end_pos - n + i  (wrapping).
-            // far_end_pos points to next-write, so most recent sample is at
-            // far_end_pos - 1.  For the i-th near-end sample we want the
-            // far-end "now" to be at (far_end_pos - n + i).  We add fl
-            // repeatedly to avoid underflow on the usize subtraction.
-            let base = (self.far_end_pos + fl * ((n / fl) + 2) + i - n) % fl;
-
             for k in 0..fl {
                 let fe_idx = (base + fl - k) % fl;
-                let fe = self.far_end_buf[fe_idx];
-                echo_est += self.filter_coeffs[k] * fe;
+                let fe = self.far_buf[fe_idx];
+                echo_est += self.filter[k] * fe;
                 power += fe * fe;
             }
 
             let error = near_f - echo_est;
 
-            // --- NLMS coefficient update ---
-            let norm = power + 1.0; // +1 regularisation to avoid div-by-zero
-            let step = self.mu * error / norm;
-
-            for k in 0..fl {
-                let fe_idx = (base + fl - k) % fl;
-                let fe = self.far_end_buf[fe_idx];
-                self.filter_coeffs[k] += step * fe;
+            // --- NLMS adaptation (only when far-end active & no double-talk) ---
+            if far_active && !is_doubletalk && power > 10.0 {
+                let step = self.mu * error / (power + 1.0);
+                for k in 0..fl {
+                    let fe_idx = (base + fl - k) % fl;
+                    self.filter[k] += step * self.far_buf[fe_idx];
+                }
             }
 
-            // Clamp output
-            let out = error.max(-32768.0).min(32767.0);
+            let out = error.clamp(-32768.0, 32767.0);
             nearend[i] = out as i16;
 
-            sum_near_sq += (near_f as f64) * (near_f as f64);
-            sum_err_sq += (out as f64) * (out as f64);
+            sum_near_sq += (near_f as f64).powi(2);
+            sum_err_sq += (out as f64).powi(2);
         }
 
-        // ERLE ratio
         if sum_err_sq < 1.0 {
-            return 100.0; // near-perfect cancellation
+            100.0
+        } else {
+            (sum_near_sq / sum_err_sq).sqrt() as f32
         }
-        (sum_near_sq / sum_err_sq).sqrt() as f32
     }
 
-    /// Enable or disable echo cancellation.
     pub fn set_enabled(&mut self, enabled: bool) {
         self.enabled = enabled;
     }
 
-    /// Returns whether echo cancellation is currently enabled.
     pub fn is_enabled(&self) -> bool {
         self.enabled
     }
 
-    /// Reset the adaptive filter to its initial state.
-    ///
-    /// Zeroes out all filter coefficients and the far-end circular buffer.
     pub fn reset(&mut self) {
-        self.filter_coeffs.iter_mut().for_each(|c| *c = 0.0);
-        self.far_end_buf.iter_mut().for_each(|s| *s = 0.0);
-        self.far_end_pos = 0;
+        self.filter.iter_mut().for_each(|c| *c = 0.0);
+        self.far_buf.iter_mut().for_each(|s| *s = 0.0);
+        self.far_pos = 0;
+        self.far_peak = 0.0;
+        self.delay_ring.iter_mut().for_each(|s| *s = 0.0);
+        self.delay_write = 0;
+        self.delay_read = 0;
+        self.dtd_holdover = 0;
     }
 }
 
@@ -143,50 +228,40 @@ mod tests {
     use super::*;
 
     #[test]
-    fn aec_creates_with_correct_filter_len() {
-        let aec = EchoCanceller::new(48000, 100);
-        assert_eq!(aec.filter_len, 4800);
-        assert_eq!(aec.filter_coeffs.len(), 4800);
-        assert_eq!(aec.far_end_buf.len(), 4800);
+    fn creates_with_correct_sizes() {
+        let aec = EchoCanceller::with_delay(48000, 60, 40);
+        assert_eq!(aec.filter_len, 2880); // 60ms @ 48kHz
+        assert_eq!(aec.delay_samples, 1920); // 40ms @ 48kHz
     }
 
     #[test]
-    fn aec_passthrough_when_disabled() {
-        let mut aec = EchoCanceller::new(48000, 100);
+    fn passthrough_when_disabled() {
+        let mut aec = EchoCanceller::new(48000, 60);
         aec.set_enabled(false);
-        assert!(!aec.is_enabled());
 
-        let original: Vec<i16> = (0..480).map(|i| (i * 10) as i16).collect();
+        let original: Vec<i16> = (0..960).map(|i| (i * 10) as i16).collect();
         let mut frame = original.clone();
-        let erle = aec.process_frame(&mut frame);
-        assert_eq!(erle, 1.0);
+        aec.process_frame(&mut frame);
         assert_eq!(frame, original);
     }
 
     #[test]
-    fn aec_reset_zeroes_state() {
-        let mut aec = EchoCanceller::new(48000, 10); // short for test speed
-        let farend: Vec<i16> = (0..480).map(|i| ((i * 37) % 1000) as i16).collect();
-        aec.feed_farend(&farend);
-
-        aec.reset();
-
-        assert!(aec.filter_coeffs.iter().all(|&c| c == 0.0));
-        assert!(aec.far_end_buf.iter().all(|&s| s == 0.0));
-        assert_eq!(aec.far_end_pos, 0);
+    fn silence_passthrough() {
+        let mut aec = EchoCanceller::with_delay(48000, 30, 0);
+        aec.feed_farend(&vec![0i16; 960]);
+        let mut frame = vec![0i16; 960];
+        aec.process_frame(&mut frame);
+        assert!(frame.iter().all(|&s| s == 0));
     }
 
     #[test]
-    fn aec_reduces_echo_of_known_signal() {
-        // Use a small filter for speed.  Feed a known far-end signal, then
-        // present the *same* signal as near-end (perfect echo, no room).
-        // After adaptation the output energy should drop.
-        let filter_ms = 5; // 240 taps at 48 kHz
-        let mut aec = EchoCanceller::new(48000, filter_ms);
+    fn reduces_echo_with_no_delay() {
+        // Simulate: far-end plays, echo arrives at mic attenuated by ~50%
+        // (realistic — speaker to mic on laptop loses volume).
+        let mut aec = EchoCanceller::with_delay(48000, 10, 0);
 
-        // Generate a simple repeating pattern.
-        let frame_len = 480usize;
-        let make_frame = |offset: usize| -> Vec<i16> {
+        let frame_len = 480;
+        let make_tone = |offset: usize| -> Vec<i16> {
             (0..frame_len)
                 .map(|i| {
                     let t = (offset + i) as f64 / 48000.0;
@@ -195,18 +270,16 @@ mod tests {
                 .collect()
         };
 
-        // Warm up the adaptive filter with several frames.
         let mut last_erle = 1.0f32;
-        for frame_idx in 0..40 {
-            let farend = make_frame(frame_idx * frame_len);
+        for frame_idx in 0..100 {
+            let farend = make_tone(frame_idx * frame_len);
             aec.feed_farend(&farend);
 
-            // Near-end = exact copy of far-end (pure echo).
-            let mut nearend = farend.clone();
+            // Near-end = attenuated copy of far-end (echo at ~50% volume).
+            let mut nearend: Vec<i16> = farend.iter().map(|&s| s / 2).collect();
             last_erle = aec.process_frame(&mut nearend);
         }
 
-        // After 40 frames the ERLE should be meaningfully > 1.
         assert!(
             last_erle > 1.0,
             "expected ERLE > 1.0 after adaptation, got {last_erle}"
@@ -214,15 +287,49 @@ mod tests {
     }
 
     #[test]
-    fn aec_silence_passthrough() {
-        let mut aec = EchoCanceller::new(48000, 10);
-        // Feed silence far-end
-        aec.feed_farend(&vec![0i16; 480]);
-        // Near-end is silence too
-        let mut frame = vec![0i16; 480];
-        let erle = aec.process_frame(&mut frame);
-        assert!(erle >= 1.0);
-        // Output should still be silence
-        assert!(frame.iter().all(|&s| s == 0));
+    fn preserves_nearend_during_doubletalk() {
+        let mut aec = EchoCanceller::with_delay(48000, 30, 0);
+
+        let frame_len = 960;
+        let nearend: Vec<i16> = (0..frame_len)
+            .map(|i| {
+                let t = i as f64 / 48000.0;
+                (10000.0 * (2.0 * std::f64::consts::PI * 440.0 * t).sin()) as i16
+            })
+            .collect();
+
+        // Feed silence as far-end (no echo source).
+        aec.feed_farend(&vec![0i16; frame_len]);
+
+        let mut frame = nearend.clone();
+        aec.process_frame(&mut frame);
+
+        let input_energy: f64 = nearend.iter().map(|&s| (s as f64).powi(2)).sum();
+        let output_energy: f64 = frame.iter().map(|&s| (s as f64).powi(2)).sum();
+        let ratio = output_energy / input_energy;
+
+        assert!(
+            ratio > 0.8,
+            "near-end speech should be preserved, energy ratio = {ratio:.3}"
+        );
+    }
+
+    #[test]
+    fn delay_buffer_holds_samples() {
+        let mut aec = EchoCanceller::with_delay(48000, 10, 20);
+        // 20ms delay = 960 samples @ 48kHz.
+        // After feeding, feed_farend auto-drains available samples to far_buf.
+        // So delay_available() is always 0 after feed_farend returns.
+        // Instead, verify far_pos advances only after the delay is filled.
+
+        // Feed 960 samples (= delay amount). No samples released yet.
+        aec.feed_farend(&vec![1i16; 960]);
+        // far_buf should still be all zeros (nothing released).
+        assert!(aec.far_buf.iter().all(|&s| s == 0.0), "nothing should be released yet");
+
+        // Feed 480 more. 480 should be released to far_buf.
+        aec.feed_farend(&vec![2i16; 480]);
+        let non_zero = aec.far_buf.iter().filter(|&&s| s != 0.0).count();
+        assert!(non_zero > 0, "samples should have been released to far_buf");
     }
 }

crates/wzp-codec/src/dred_ffi.rs

@@ -0,0 +1,585 @@
+//! Raw opusic-sys FFI wrappers for libopus 1.5.2 decoder + DRED reconstruction.
+//!
+//! # Why this module exists
+//!
+//! We cannot use `opusic_c::Decoder` because its inner `*mut OpusDecoder`
+//! pointer is `pub(crate)` — not reachable from outside the opusic-c crate.
+//! Phase 3 of the DRED integration needs to hand that same pointer to
+//! `opus_decoder_dred_decode`, and running two parallel decoders (one from
+//! opusic-c for normal audio, another from opusic-sys for DRED) would cause
+//! the DRED-only decoder's internal state to drift out of sync with the
+//! audio stream because it would not see normal decode calls.
+//!
+//! The fix is to own the raw decoder ourselves and use the same handle for
+//! both normal decode AND DRED reconstruction. This module is the single
+//! owner of `*mut OpusDecoder`, `*mut OpusDREDDecoder`, and `*mut OpusDRED`
+//! in the WZP workspace.
+//!
+//! # Phase 3a scope
+//!
+//! Phase 0 added `DecoderHandle` (normal decode). Phase 3a adds:
+//! - [`DredDecoderHandle`] — wraps `*mut OpusDREDDecoder` for parsing DRED
+//!   side-channel data out of arriving Opus packets.
+//! - [`DredState`] — wraps `*mut OpusDRED` (a fixed 10,592-byte buffer
+//!   allocated by libopus) that holds parsed DRED state between the parse
+//!   and reconstruct steps.
+//! - [`DredDecoderHandle::parse_into`] — wraps `opus_dred_parse`.
+//! - [`DecoderHandle::reconstruct_from_dred`] — wraps `opus_decoder_dred_decode`.
+//!
+//! The pattern is: on every arriving Opus packet, the receiver calls
+//! `parse_into` with a reusable `DredState`, then stores (seq, state_clone)
+//! in a ring. On detected loss, the receiver computes the offset from the
+//! freshest reachable DRED state and calls `reconstruct_from_dred` to
+//! synthesize the missing audio.
+
+use std::ptr::NonNull;
+
+use opusic_sys::{
+    OPUS_OK, OpusDRED, OpusDREDDecoder, OpusDecoder as RawOpusDecoder, opus_decode,
+    opus_decoder_create, opus_decoder_destroy, opus_decoder_dred_decode, opus_dred_alloc,
+    opus_dred_decoder_create, opus_dred_decoder_destroy, opus_dred_free, opus_dred_parse,
+};
+use wzp_proto::CodecError;
+
+/// libopus operates at 48 kHz for all Opus variants we use.
+const SAMPLE_RATE_HZ: i32 = 48_000;
+/// Mono.
+const CHANNELS: i32 = 1;
+
+/// Safe owner of a `*mut OpusDecoder` allocated via `opus_decoder_create`.
+///
+/// Releases the decoder in `Drop`. All FFI access goes through `&mut self`
+/// methods, so there is no aliasing or race. The raw pointer is exposed via
+/// [`Self::as_raw_ptr`] at a crate-internal visibility for the future Phase 3
+/// DRED reconstruction path — external crates cannot reach it.
+pub struct DecoderHandle {
+    inner: NonNull<RawOpusDecoder>,
+}
+
+impl DecoderHandle {
+    /// Allocate a new Opus decoder at 48 kHz mono.
+    pub fn new() -> Result<Self, CodecError> {
+        let mut error: i32 = OPUS_OK;
+        // SAFETY: opus_decoder_create writes to `error` and returns either a
+        // valid heap pointer or null. We check both before constructing the
+        // NonNull wrapper.
+        let ptr = unsafe { opus_decoder_create(SAMPLE_RATE_HZ, CHANNELS, &mut error) };
+        if error != OPUS_OK {
+            // Even if ptr is non-null on error, libopus contracts guarantee
+            // it is unusable — do not attempt to free it.
+            return Err(CodecError::DecodeFailed(format!(
+                "opus_decoder_create failed: err={error}"
+            )));
+        }
+        let inner = NonNull::new(ptr).ok_or_else(|| {
+            CodecError::DecodeFailed("opus_decoder_create returned null".into())
+        })?;
+        Ok(Self { inner })
+    }
+
+    /// Decode an Opus packet into PCM samples.
+    ///
+    /// `pcm` must have enough capacity for the frame (960 for 20 ms, 1920
+    /// for 40 ms at 48 kHz mono). Returns the number of decoded samples
+    /// per channel — for mono streams this equals the total sample count.
+    pub fn decode(&mut self, packet: &[u8], pcm: &mut [i16]) -> Result<usize, CodecError> {
+        if packet.is_empty() {
+            return Err(CodecError::DecodeFailed("empty packet".into()));
+        }
+        if pcm.is_empty() {
+            return Err(CodecError::DecodeFailed("empty output buffer".into()));
+        }
+        // SAFETY: self.inner is a valid *mut OpusDecoder owned by this struct.
+        // `data` / `pcm` are live Rust slices, so their pointers and lengths
+        // are valid for the duration of the call. libopus reads len bytes
+        // from data and writes up to frame_size samples (per channel) to pcm.
+        let n = unsafe {
+            opus_decode(
+                self.inner.as_ptr(),
+                packet.as_ptr(),
+                packet.len() as i32,
+                pcm.as_mut_ptr(),
+                pcm.len() as i32,
+                /* decode_fec = */ 0,
+            )
+        };
+        if n < 0 {
+            return Err(CodecError::DecodeFailed(format!(
+                "opus_decode failed: err={n}"
+            )));
+        }
+        Ok(n as usize)
+    }
+
+    /// Generate packet-loss concealment audio for a missing frame.
+    ///
+    /// Implemented via `opus_decode` with a null data pointer, per the
+    /// libopus API contract. `pcm` should be sized for the expected frame.
+    pub fn decode_lost(&mut self, pcm: &mut [i16]) -> Result<usize, CodecError> {
+        if pcm.is_empty() {
+            return Err(CodecError::DecodeFailed("empty output buffer".into()));
+        }
+        // SAFETY: same invariants as decode(). libopus documents that passing
+        // a null data pointer with len=0 triggers PLC synthesis into pcm.
+        let n = unsafe {
+            opus_decode(
+                self.inner.as_ptr(),
+                std::ptr::null(),
+                0,
+                pcm.as_mut_ptr(),
+                pcm.len() as i32,
+                /* decode_fec = */ 0,
+            )
+        };
+        if n < 0 {
+            return Err(CodecError::DecodeFailed(format!(
+                "opus_decode PLC failed: err={n}"
+            )));
+        }
+        Ok(n as usize)
+    }
+
+    /// Reconstruct audio from a `DredState` into the `output` buffer.
+    ///
+    /// `offset_samples` is the sample position (positive, measured backward
+    /// from the packet anchor that produced `state`) where reconstruction
+    /// begins. `output.len()` must match the number of samples to synthesize.
+    ///
+    /// The libopus API: `opus_decoder_dred_decode(st, dred, dred_offset, pcm,
+    /// frame_size)` where `dred_offset` is "position of the redundancy to
+    /// decode, in samples before the beginning of the real audio data in the
+    /// packet." Valid values: `0 < offset_samples < state.samples_available()`.
+    ///
+    /// Returns the number of samples actually written (should equal
+    /// `output.len()` on success).
+    pub fn reconstruct_from_dred(
+        &mut self,
+        state: &DredState,
+        offset_samples: i32,
+        output: &mut [i16],
+    ) -> Result<usize, CodecError> {
+        if output.is_empty() {
+            return Err(CodecError::DecodeFailed(
+                "empty reconstruction output buffer".into(),
+            ));
+        }
+        if offset_samples <= 0 {
+            return Err(CodecError::DecodeFailed(format!(
+                "DRED offset must be positive (got {offset_samples})"
+            )));
+        }
+        if offset_samples > state.samples_available() {
+            return Err(CodecError::DecodeFailed(format!(
+                "DRED offset {offset_samples} exceeds available samples {}",
+                state.samples_available()
+            )));
+        }
+        // SAFETY: self.inner is a valid *mut OpusDecoder, state.inner is a
+        // valid *const OpusDRED populated by a prior parse_into call, and
+        // output is a live mutable slice. libopus reads from dred and writes
+        // exactly frame_size samples (the output.len()) to pcm.
+        let n = unsafe {
+            opus_decoder_dred_decode(
+                self.inner.as_ptr(),
+                state.inner.as_ptr(),
+                offset_samples,
+                output.as_mut_ptr(),
+                output.len() as i32,
+            )
+        };
+        if n < 0 {
+            return Err(CodecError::DecodeFailed(format!(
+                "opus_decoder_dred_decode failed: err={n}"
+            )));
+        }
+        Ok(n as usize)
+    }
+}
+
+impl Drop for DecoderHandle {
+    fn drop(&mut self) {
+        // SAFETY: we own the pointer and no further access happens after
+        // this call because Drop consumes self.
+        unsafe { opus_decoder_destroy(self.inner.as_ptr()) };
+    }
+}
+
+// SAFETY: The underlying OpusDecoder is a plain heap allocation with no
+// thread-local or lock-free state. It is safe to move between threads
+// (Send), and all method access is gated by &mut self so Rust's borrow
+// checker prevents simultaneous access from multiple threads (Sync).
+unsafe impl Send for DecoderHandle {}
+unsafe impl Sync for DecoderHandle {}
+
+// ─── DRED decoder (parser) ──────────────────────────────────────────────────
+
+/// Safe owner of a `*mut OpusDREDDecoder` allocated via
+/// `opus_dred_decoder_create`.
+///
+/// The DRED decoder is a **separate** libopus object from the regular
+/// `OpusDecoder`. It's used exclusively for parsing DRED side-channel data
+/// out of arriving Opus packets via [`Self::parse_into`]. Actual audio
+/// reconstruction from the parsed state uses the regular `DecoderHandle`
+/// via [`DecoderHandle::reconstruct_from_dred`].
+pub struct DredDecoderHandle {
+    inner: NonNull<OpusDREDDecoder>,
+}
+
+impl DredDecoderHandle {
+    /// Allocate a new DRED decoder.
+    pub fn new() -> Result<Self, CodecError> {
+        let mut error: i32 = OPUS_OK;
+        // SAFETY: opus_dred_decoder_create writes to `error` and returns
+        // either a valid heap pointer or null. Both are checked.
+        let ptr = unsafe { opus_dred_decoder_create(&mut error) };
+        if error != OPUS_OK {
+            return Err(CodecError::DecodeFailed(format!(
+                "opus_dred_decoder_create failed: err={error}"
+            )));
+        }
+        let inner = NonNull::new(ptr).ok_or_else(|| {
+            CodecError::DecodeFailed("opus_dred_decoder_create returned null".into())
+        })?;
+        Ok(Self { inner })
+    }
+
+    /// Parse DRED side-channel data from an Opus packet into `state`.
+    ///
+    /// Returns the number of samples of audio history available for
+    /// reconstruction, or 0 if the packet carries no DRED data. Subsequent
+    /// `DecoderHandle::reconstruct_from_dred` calls using this `state` can
+    /// reconstruct any sample position in `(0, samples_available]`.
+    ///
+    /// libopus API: `opus_dred_parse(dred_dec, dred, data, len,
+    /// max_dred_samples, sampling_rate, dred_end, defer_processing)`. We
+    /// pass `max_dred_samples = 48000` (1 s at 48 kHz, the DRED maximum),
+    /// `sampling_rate = 48000`, `defer_processing = 0` (process immediately).
+    /// The `dred_end` output is the silence gap at the tail of the DRED
+    /// window; we subtract it from the total offset to give callers the
+    /// truly usable sample count.
+    pub fn parse_into(
+        &mut self,
+        state: &mut DredState,
+        packet: &[u8],
+    ) -> Result<i32, CodecError> {
+        if packet.is_empty() {
+            state.samples_available = 0;
+            return Ok(0);
+        }
+        let mut dred_end: i32 = 0;
+        // SAFETY: self.inner is a valid *mut OpusDREDDecoder; state.inner is
+        // a valid *mut OpusDRED allocated via opus_dred_alloc; packet is a
+        // live slice; dred_end is a stack int. libopus reads packet bytes
+        // and writes parsed DRED state into *state.inner.
+        let ret = unsafe {
+            opus_dred_parse(
+                self.inner.as_ptr(),
+                state.inner.as_ptr(),
+                packet.as_ptr(),
+                packet.len() as i32,
+                /* max_dred_samples = */ 48_000, // 1s max per libopus 1.5
+                /* sampling_rate = */ 48_000,
+                &mut dred_end,
+                /* defer_processing = */ 0,
+            )
+        };
+        if ret < 0 {
+            state.samples_available = 0;
+            return Err(CodecError::DecodeFailed(format!(
+                "opus_dred_parse failed: err={ret}"
+            )));
+        }
+        // ret is the positive offset of the first decodable DRED sample,
+        // or 0 if no DRED is present. dred_end is the silence gap at the
+        // tail. The usable sample range is (dred_end, ret], so the count
+        // of usable samples is ret - dred_end. We store `ret` as the max
+        // usable offset — callers should pass dred_offset values in the
+        // range (dred_end, ret] to reconstruct_from_dred. For simplicity
+        // we expose just samples_available = ret and let callers treat
+        // the full window as valid (the silence gap is small and libopus
+        // handles minor boundary cases gracefully).
+        state.samples_available = ret;
+        Ok(ret)
+    }
+}
+
+impl Drop for DredDecoderHandle {
+    fn drop(&mut self) {
+        // SAFETY: we own the pointer and no further access happens after
+        // this call because Drop consumes self.
+        unsafe { opus_dred_decoder_destroy(self.inner.as_ptr()) };
+    }
+}
+
+// SAFETY: same reasoning as DecoderHandle — heap allocation with no
+// thread-local state, &mut self access discipline prevents races.
+unsafe impl Send for DredDecoderHandle {}
+unsafe impl Sync for DredDecoderHandle {}
+
+// ─── DRED state buffer ──────────────────────────────────────────────────────
+
+/// Safe owner of a `*mut OpusDRED` allocated via `opus_dred_alloc`.
+///
+/// Holds a fixed-size (10,592-byte per libopus 1.5) buffer that
+/// `DredDecoderHandle::parse_into` populates from an Opus packet. The state
+/// is reusable — the caller can call `parse_into` again on the same
+/// `DredState` to overwrite it with a fresh packet's data.
+///
+/// `samples_available` tracks the last-parsed result so reconstruction
+/// callers don't need to thread the return value separately. A fresh
+/// state (before any `parse_into`) has `samples_available == 0`.
+pub struct DredState {
+    inner: NonNull<OpusDRED>,
+    samples_available: i32,
+}
+
+impl DredState {
+    /// Allocate a new DRED state buffer.
+    pub fn new() -> Result<Self, CodecError> {
+        let mut error: i32 = OPUS_OK;
+        // SAFETY: opus_dred_alloc writes to `error` and returns either a
+        // valid heap pointer or null.
+        let ptr = unsafe { opus_dred_alloc(&mut error) };
+        if error != OPUS_OK {
+            return Err(CodecError::DecodeFailed(format!(
+                "opus_dred_alloc failed: err={error}"
+            )));
+        }
+        let inner = NonNull::new(ptr)
+            .ok_or_else(|| CodecError::DecodeFailed("opus_dred_alloc returned null".into()))?;
+        Ok(Self {
+            inner,
+            samples_available: 0,
+        })
+    }
+
+    /// How many samples of audio history this state currently covers.
+    ///
+    /// Returns 0 if the state is fresh or the last parse found no DRED
+    /// data. Otherwise returns the positive offset set by the most recent
+    /// `DredDecoderHandle::parse_into` call — the maximum valid
+    /// `offset_samples` value for `DecoderHandle::reconstruct_from_dred`.
+    pub fn samples_available(&self) -> i32 {
+        self.samples_available
+    }
+
+    /// Reset the state to "fresh" without freeing the underlying buffer.
+    /// The next `parse_into` will overwrite the contents.
+    pub fn reset(&mut self) {
+        self.samples_available = 0;
+    }
+}
+
+impl Drop for DredState {
+    fn drop(&mut self) {
+        // SAFETY: we own the pointer and no further access happens after
+        // this call because Drop consumes self.
+        unsafe { opus_dred_free(self.inner.as_ptr()) };
+    }
+}
+
+// SAFETY: same reasoning as DecoderHandle.
+unsafe impl Send for DredState {}
+unsafe impl Sync for DredState {}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn decoder_handle_creates_and_drops() {
+        let handle = DecoderHandle::new().expect("decoder create");
+        // Dropping the handle must not panic or leak — validated by miri
+        // and the absence of sanitizer complaints in CI.
+        drop(handle);
+    }
+
+    #[test]
+    fn decode_lost_produces_full_frame_of_silence_on_cold_start() {
+        let mut handle = DecoderHandle::new().unwrap();
+        // 20 ms @ 48 kHz mono.
+        let mut pcm = vec![0i16; 960];
+        let n = handle.decode_lost(&mut pcm).unwrap();
+        assert_eq!(n, 960);
+        // On a fresh decoder, PLC output is silence (no past audio to extend).
+        assert!(pcm.iter().all(|&s| s == 0));
+    }
+
+    #[test]
+    fn decode_empty_packet_errors() {
+        let mut handle = DecoderHandle::new().unwrap();
+        let mut pcm = vec![0i16; 960];
+        let err = handle.decode(&[], &mut pcm);
+        assert!(err.is_err());
+    }
+
+    // ─── Phase 3a — DRED decoder + state ────────────────────────────────────
+
+    #[test]
+    fn dred_decoder_handle_creates_and_drops() {
+        let h = DredDecoderHandle::new().expect("dred decoder create");
+        drop(h);
+    }
+
+    #[test]
+    fn dred_state_creates_and_drops() {
+        let s = DredState::new().expect("dred state alloc");
+        assert_eq!(s.samples_available(), 0);
+        drop(s);
+    }
+
+    #[test]
+    fn dred_state_reset_zeroes_counter() {
+        let mut s = DredState::new().unwrap();
+        s.samples_available = 480; // pretend a parse populated it
+        assert_eq!(s.samples_available(), 480);
+        s.reset();
+        assert_eq!(s.samples_available(), 0);
+    }
+
+    /// Phase 3a end-to-end: encode a DRED-enabled stream, parse state out
+    /// of packets, and reconstruct audio at a past offset. Validates the
+    /// full parse → reconstruct pipeline against a real libopus 1.5.2
+    /// encoder so we catch FFI-layer bugs early.
+    #[test]
+    fn dred_parse_and_reconstruct_roundtrip() {
+        use crate::opus_enc::OpusEncoder;
+        use wzp_proto::{AudioEncoder, QualityProfile};
+
+        // Encoder with DRED at Opus 24k / 200 ms duration (Phase 1 default
+        // for GOOD profile). The loss floor is 5% per Phase 1.
+        let mut enc = OpusEncoder::new(QualityProfile::GOOD).unwrap();
+
+        // Decode-side handles.
+        let mut dec = DecoderHandle::new().unwrap();
+        let mut dred_dec = DredDecoderHandle::new().unwrap();
+        let mut state = DredState::new().unwrap();
+
+        // Generate 60 frames (1.2 s) of a voice-like 300 Hz sine wave so
+        // the encoder's DRED emitter has real content to encode rather
+        // than compressing silence.
+        let frame_len = 960usize; // 20 ms @ 48 kHz
+        let make_frame = |offset: usize| -> Vec<i16> {
+            (0..frame_len)
+                .map(|i| {
+                    let t = (offset + i) as f64 / 48_000.0;
+                    (8000.0 * (2.0 * std::f64::consts::PI * 300.0 * t).sin()) as i16
+                })
+                .collect()
+        };
+
+        // Track the freshest packet that carried non-zero DRED state.
+        let mut best_samples_available = 0;
+        let mut best_packet: Option<Vec<u8>> = None;
+
+        for frame_idx in 0..60 {
+            let pcm = make_frame(frame_idx * frame_len);
+            let mut encoded = vec![0u8; 512];
+            let n = enc.encode(&pcm, &mut encoded).unwrap();
+            encoded.truncate(n);
+
+            // Run the packet through the normal decode path so dec's
+            // internal state mirrors the full stream — this is necessary
+            // for DRED reconstruction to produce meaningful output.
+            let mut decoded = vec![0i16; frame_len];
+            dec.decode(&encoded, &mut decoded).unwrap();
+
+            // Parse DRED state out of the same packet. Early packets may
+            // have samples_available == 0 while the DRED encoder warms up;
+            // later packets should carry the full window.
+            match dred_dec.parse_into(&mut state, &encoded) {
+                Ok(available) => {
+                    if available > best_samples_available {
+                        best_samples_available = available;
+                        best_packet = Some(encoded.clone());
+                    }
+                }
+                Err(e) => panic!("parse_into errored unexpectedly: {e:?}"),
+            }
+        }
+
+        // By the time we're 60 frames in, DRED should have emitted data.
+        assert!(
+            best_samples_available > 0,
+            "DRED emitted zero samples across 60 frames — the encoder isn't \
+             producing DRED bytes (check set_dred_duration and packet_loss floor)"
+        );
+
+        // Parse the best packet into a fresh state and reconstruct some
+        // audio from somewhere inside its DRED window. We use frame_len/2
+        // as the offset to pick a point squarely inside the reconstructable
+        // range rather than at an edge.
+        let packet = best_packet.expect("at least one packet had DRED state");
+        let mut fresh_state = DredState::new().unwrap();
+        let available = dred_dec.parse_into(&mut fresh_state, &packet).unwrap();
+        assert!(available > 0, "re-parse of known-good packet returned 0");
+
+        // Need a decoder that's in the right state to reconstruct — rewind
+        // by creating a fresh one and feeding it the same stream up to the
+        // point of the best packet. Simpler: just use a fresh decoder and
+        // accept that the reconstructed samples may not be phase-matched.
+        // The test here only asserts *non-silent energy*, not signal fidelity.
+        let mut recon_dec = DecoderHandle::new().unwrap();
+        // Warm up the decoder with one frame so its internal state is valid.
+        let warmup_pcm = vec![0i16; frame_len];
+        let warmup_encoded = {
+            let mut warmup_enc = OpusEncoder::new(QualityProfile::GOOD).unwrap();
+            let mut buf = vec![0u8; 512];
+            let n = warmup_enc.encode(&warmup_pcm, &mut buf).unwrap();
+            buf.truncate(n);
+            buf
+        };
+        let mut throwaway = vec![0i16; frame_len];
+        let _ = recon_dec.decode(&warmup_encoded, &mut throwaway);
+
+        // Reconstruct 20 ms from some position inside the DRED window.
+        let offset = (available / 2).max(480).min(available);
+        let mut recon_pcm = vec![0i16; frame_len];
+        let n = recon_dec
+            .reconstruct_from_dred(&fresh_state, offset, &mut recon_pcm)
+            .expect("reconstruct_from_dred failed");
+        assert_eq!(n, frame_len);
+
+        // Energy check: reconstructed audio should not be all zeros. A
+        // loose threshold — the DRED reconstruction won't be phase-matched
+        // to our sine wave because we fed a cold decoder only one warmup
+        // frame, but it should still produce non-silent speech-like output
+        // since the DRED state was parsed from real speech content.
+        let energy: u64 = recon_pcm.iter().map(|&s| (s as i32).unsigned_abs() as u64).sum();
+        assert!(
+            energy > 0,
+            "reconstructed audio has zero total energy — DRED reconstruction produced silence"
+        );
+    }
+
+    /// A second roundtrip variant: offset too large errors cleanly rather
+    /// than crashing the FFI.
+    #[test]
+    fn reconstruct_with_out_of_range_offset_errors() {
+        let mut dec = DecoderHandle::new().unwrap();
+        let state = DredState::new().unwrap();
+        // state has samples_available == 0 (fresh), so any positive offset
+        // should be out of range.
+        let mut out = vec![0i16; 960];
+        let err = dec.reconstruct_from_dred(&state, 480, &mut out);
+        assert!(err.is_err());
+    }
+
+    #[test]
+    fn reconstruct_with_zero_offset_errors() {
+        let mut dec = DecoderHandle::new().unwrap();
+        let state = DredState::new().unwrap();
+        let mut out = vec![0i16; 960];
+        let err = dec.reconstruct_from_dred(&state, 0, &mut out);
+        assert!(err.is_err());
+    }
+
+    #[test]
+    fn dred_parse_empty_packet_returns_zero() {
+        let mut dred_dec = DredDecoderHandle::new().unwrap();
+        let mut state = DredState::new().unwrap();
+        let result = dred_dec.parse_into(&mut state, &[]).unwrap();
+        assert_eq!(result, 0);
+        assert_eq!(state.samples_available(), 0);
+    }
+}

crates/wzp-codec/src/lib.rs

@@ -15,6 +15,7 @@ pub mod agc;
 pub mod codec2_dec;
 pub mod codec2_enc;
 pub mod denoise;
+pub mod dred_ffi;
 pub mod opus_dec;
 pub mod opus_enc;
 pub mod resample;
@@ -27,6 +28,26 @@ pub use denoise::NoiseSupressor;
 pub use silence::{ComfortNoise, SilenceDetector};
 pub use wzp_proto::{AudioDecoder, AudioEncoder, CodecId, QualityProfile};
 
+use std::sync::atomic::{AtomicBool, Ordering};
+
+/// Global verbose-logging flag for DRED. Off by default — when enabled
+/// (via the GUI debug toggle wired through Tauri), the encoder logs its
+/// DRED config + libopus version, and the recv path logs every DRED
+/// reconstruction, classical PLC fill, and parse heartbeat. Off in
+/// "normal" mode keeps logcat clean.
+static DRED_VERBOSE_LOGS: AtomicBool = AtomicBool::new(false);
+
+/// Returns whether DRED verbose logging is currently enabled.
+#[inline]
+pub fn dred_verbose_logs() -> bool {
+    DRED_VERBOSE_LOGS.load(Ordering::Relaxed)
+}
+
+/// Enable/disable DRED verbose logging at runtime.
+pub fn set_dred_verbose_logs(enabled: bool) {
+    DRED_VERBOSE_LOGS.store(enabled, Ordering::Relaxed);
+}
+
 /// Create an adaptive encoder starting at the given quality profile.
 ///
 /// The returned encoder accepts 48 kHz mono PCM regardless of the active

crates/wzp-codec/src/opus_dec.rs

@@ -1,30 +1,32 @@
-//! Opus decoder wrapping the `audiopus` crate.
+//! Opus decoder built on top of the raw opusic-sys `DecoderHandle`.
+//!
+//! Phase 0 of the DRED integration: we went straight to a custom
+//! `DecoderHandle` instead of `opusic_c::Decoder` because the latter's
+//! inner pointer is `pub(crate)` and we need to reach it in Phase 3 for
+//! `opus_decoder_dred_decode`. See `dred_ffi.rs` for the rationale and
+//! `docs/PRD-dred-integration.md` for the full plan.
 
-use audiopus::coder::Decoder;
-use audiopus::{Channels, MutSignals, SampleRate};
-use audiopus::packet::Packet;
+use crate::dred_ffi::{DecoderHandle, DredState};
 use wzp_proto::{AudioDecoder, CodecError, CodecId, QualityProfile};
 
-/// Opus decoder implementing `AudioDecoder`.
+/// Opus decoder implementing [`AudioDecoder`].
 ///
-/// Operates at 48 kHz mono output.
+/// Operates at 48 kHz mono output. 20 ms and 40 ms frames supported via
+/// the active `QualityProfile`. Behavior is intentionally identical to
+/// the pre-swap audiopus-based decoder at this phase — DRED reconstruction
+/// lands in Phase 3.
 pub struct OpusDecoder {
-    inner: Decoder,
+    inner: DecoderHandle,
     codec_id: CodecId,
     frame_duration_ms: u8,
 }
 
-// SAFETY: Same reasoning as OpusEncoder — exclusive access via &mut self.
-unsafe impl Sync for OpusDecoder {}
-
 impl OpusDecoder {
     /// Create a new Opus decoder for the given quality profile.
     pub fn new(profile: QualityProfile) -> Result<Self, CodecError> {
-        let decoder = Decoder::new(SampleRate::Hz48000, Channels::Mono)
-            .map_err(|e| CodecError::DecodeFailed(format!("opus decoder init: {e}")))?;
-
+        let inner = DecoderHandle::new()?;
         Ok(Self {
-            inner: decoder,
+            inner,
             codec_id: profile.codec,
             frame_duration_ms: profile.frame_duration_ms,
         })
@@ -34,6 +36,24 @@ impl OpusDecoder {
     pub fn frame_samples(&self) -> usize {
         (48_000 * self.frame_duration_ms as usize) / 1000
     }
+
+    /// Reconstruct a lost frame from a previously parsed `DredState`.
+    ///
+    /// Phase 3b entry point: callers (CallDecoder / engine.rs) use this to
+    /// synthesize audio for gaps detected by the jitter buffer when DRED
+    /// side-channel state from a later-arriving packet covers the gap's
+    /// sample offset. `offset_samples` is measured backward from the anchor
+    /// packet that produced `state`. See `DecoderHandle::reconstruct_from_dred`
+    /// for the full semantics.
+    pub fn reconstruct_from_dred(
+        &mut self,
+        state: &DredState,
+        offset_samples: i32,
+        output: &mut [i16],
+    ) -> Result<usize, CodecError> {
+        self.inner
+            .reconstruct_from_dred(state, offset_samples, output)
+    }
 }
 
 impl AudioDecoder for OpusDecoder {
@@ -45,15 +65,7 @@ impl AudioDecoder for OpusDecoder {
                 pcm.len()
             )));
         }
-        let packet = Packet::try_from(encoded)
-            .map_err(|e| CodecError::DecodeFailed(format!("invalid packet: {e}")))?;
-        let signals = MutSignals::try_from(pcm)
-            .map_err(|e| CodecError::DecodeFailed(format!("output signals: {e}")))?;
-        let n = self
-            .inner
-            .decode(Some(packet), signals, false)
-            .map_err(|e| CodecError::DecodeFailed(format!("opus decode: {e}")))?;
-        Ok(n)
+        self.inner.decode(encoded, pcm)
     }
 
     fn decode_lost(&mut self, pcm: &mut [i16]) -> Result<usize, CodecError> {
@@ -64,13 +76,7 @@ impl AudioDecoder for OpusDecoder {
                 pcm.len()
             )));
         }
-        let signals = MutSignals::try_from(pcm)
-            .map_err(|e| CodecError::DecodeFailed(format!("output signals: {e}")))?;
-        let n = self
-            .inner
-            .decode(None, signals, false)
-            .map_err(|e| CodecError::DecodeFailed(format!("opus PLC: {e}")))?;
-        Ok(n)
+        self.inner.decode_lost(pcm)
     }
 
     fn codec_id(&self) -> CodecId {

crates/wzp-codec/src/opus_enc.rs

@@ -1,58 +1,225 @@
-//! Opus encoder wrapping the `audiopus` crate.
+//! Opus encoder wrapping the `opusic-c` crate (libopus 1.5.2).
+//!
+//! Phase 1 of the DRED integration: encoder-side DRED is enabled on every
+//! Opus profile with a tiered duration (studio 100 ms / normal 200 ms /
+//! degraded 500 ms), and Opus inband FEC (LBRR) is disabled because DRED
+//! is the stronger mechanism for the same failure mode. The legacy behavior
+//! is preserved behind the `AUDIO_USE_LEGACY_FEC` environment variable as a
+//! runtime escape hatch for rollout. See `docs/PRD-dred-integration.md`.
+//!
+//! # DRED duration policy
+//!
+//! Rationale from the PRD:
+//! - Studio tiers (Opus 32k/48k/64k): 100 ms — loss is rare on high-quality
+//!   networks; short window keeps decoder CPU modest.
+//! - Normal tiers (Opus 16k/24k): 200 ms — balanced baseline covering common
+//!   VoIP loss patterns (20–150 ms bursts from wifi roam, transient congestion).
+//! - Degraded tier (Opus 6k): 1040 ms — users on 6k are by definition on a
+//!   bad link; the maximum libopus DRED window buys the best burst resilience
+//!   where it matters. The RDO-VAE naturally degrades quality at longer offsets.
+//!
+//! # Why the 15% packet loss floor
+//!
+//! libopus 1.5's DRED emitter is gated on `OPUS_SET_PACKET_LOSS_PERC` and
+//! scales the emitted window proportionally to the assumed loss:
+//!
+//! ```text
+//!   loss_pct  samples_available    effective_ms
+//!   5%         720                   15
+//!   10%        2640                  55
+//!   15%        4560                  95
+//!   20%        6480                 135
+//!   25%+       8400 (capped)        175  (≈ 87% of the 200ms configured max)
+//! ```
+//!
+//! Measured empirically against libopus 1.5.2 on Opus 24k / 200 ms DRED
+//! duration during Phase 3b. At 5% loss the window is only 15 ms — too
+//! small to even reconstruct a single 20 ms Opus frame. 15% gives 95 ms
+//! (enough for single-frame recovery plus modest burst margin) while
+//! keeping the bitrate overhead modest compared to 25%. Real measurements
+//! from the quality adapter override upward when loss exceeds the floor.
 
-use audiopus::coder::Encoder;
-use audiopus::{Application, Bitrate, Channels, SampleRate, Signal};
-use tracing::debug;
+use std::sync::OnceLock;
+
+use opusic_c::{Application, Bitrate, Channels, Encoder, InbandFec, SampleRate, Signal};
+use tracing::{debug, info, warn};
 use wzp_proto::{AudioEncoder, CodecError, CodecId, QualityProfile};
 
+/// Logged exactly once per process the first time an OpusEncoder is built.
+/// Confirms that libopus 1.5.2 (the version with DRED) is actually linked
+/// at runtime — invaluable when chasing "is the new codec loaded?"
+/// regressions on Android, where the only debug surface is logcat.
+static LIBOPUS_VERSION_LOGGED: OnceLock<()> = OnceLock::new();
+
+/// Minimum `OPUS_SET_PACKET_LOSS_PERC` value used in DRED mode. libopus
+/// scales the DRED emission window with the assumed loss percentage:
+/// empirically, 5% gives a 15 ms window (useless), 10% gives 55 ms, 15%
+/// gives 95 ms, and 25%+ saturates the configured max (~175 ms at 200 ms
+/// duration). 15% is the minimum value that produces a DRED window larger
+/// than a single 20 ms frame, making it the minimum floor that actually
+/// gives DRED something useful to reconstruct. Real loss measurements from
+/// the quality adapter override this upward.
+const DRED_LOSS_FLOOR_PCT: u8 = 15;
+
+/// Environment variable that reverts Phase 1 behavior to Phase 0 (inband FEC
+/// on, DRED off, no loss floor). Read once per encoder construction.
+const LEGACY_FEC_ENV: &str = "AUDIO_USE_LEGACY_FEC";
+
+/// Returns the DRED duration in 10 ms frame units for a given Opus codec.
+///
+/// Unit: each frame is 10 ms, so the max value of 104 corresponds to 1040 ms
+/// of reconstructable history. Returns 0 for non-Opus codecs (DRED is not
+/// emitted by the libopus encoder in that case anyway, but we avoid a
+/// pointless FFI call).
+///
+/// See the DRED duration policy in the module docs for per-tier rationale.
+pub fn dred_duration_for(codec: CodecId) -> u8 {
+    match codec {
+        // Studio tiers — loss is rare, short window.
+        CodecId::Opus32k | CodecId::Opus48k | CodecId::Opus64k => 10,
+        // Normal tiers — balanced baseline.
+        CodecId::Opus16k | CodecId::Opus24k => 20,
+        // Degraded tier — maximum burst resilience. 104 × 10 ms = 1040 ms,
+        // the highest value libopus 1.5 supports. Users on 6k are on a bad
+        // link by definition; the RDO-VAE naturally degrades quality at longer
+        // offsets, so the extra window costs only ~1-2 kbps additional overhead
+        // while buying substantially better burst resilience (up from 500 ms).
+        CodecId::Opus6k => 104,
+        // Non-Opus (Codec2 / CN): DRED is N/A.
+        CodecId::Codec2_1200 | CodecId::Codec2_3200 | CodecId::ComfortNoise => 0,
+    }
+}
+
+/// Returns whether the legacy-FEC escape hatch is active.
+///
+/// Read from `AUDIO_USE_LEGACY_FEC`. Any non-empty value activates legacy
+/// mode; unset or empty leaves DRED enabled.
+fn read_legacy_fec_env() -> bool {
+    match std::env::var(LEGACY_FEC_ENV) {
+        Ok(v) => !v.is_empty() && v != "0" && v.to_ascii_lowercase() != "false",
+        Err(_) => false,
+    }
+}
+
 /// Opus encoder implementing `AudioEncoder`.
 ///
-/// Operates at 48 kHz mono. Supports frame sizes of 20 ms (960 samples)
-/// and 40 ms (1920 samples).
+/// Operates at 48 kHz mono. Supports 20 ms and 40 ms frames via the active
+/// `QualityProfile`.
 pub struct OpusEncoder {
     inner: Encoder,
     codec_id: CodecId,
     frame_duration_ms: u8,
+    /// When `true`, revert to the Phase 0 behavior: inband FEC Mode1, DRED
+    /// disabled, no loss floor. Captured at construction time and not
+    /// re-read mid-call.
+    legacy_fec_mode: bool,
 }
 
 // SAFETY: OpusEncoder is only used via `&mut self` methods. The inner
-// audiopus Encoder contains a raw pointer that is !Sync, but we never
-// share it across threads without exclusive access.
+// opusic-c Encoder wraps a non-null pointer that is !Sync by default,
+// but we never share it across threads without exclusive access.
 unsafe impl Sync for OpusEncoder {}
 
 impl OpusEncoder {
     /// Create a new Opus encoder for the given quality profile.
     pub fn new(profile: QualityProfile) -> Result<Self, CodecError> {
-        let encoder = Encoder::new(SampleRate::Hz48000, Channels::Mono, Application::Voip)
-            .map_err(|e| CodecError::EncodeFailed(format!("opus encoder init: {e}")))?;
+        // opusic-c argument order: (Channels, SampleRate, Application)
+        // — different from audiopus's (SampleRate, Channels, Application).
+        let encoder = Encoder::new(Channels::Mono, SampleRate::Hz48000, Application::Voip)
+            .map_err(|e| CodecError::EncodeFailed(format!("opus encoder init: {e:?}")))?;
+
+        let legacy_fec_mode = read_legacy_fec_env();
+        if legacy_fec_mode {
+            warn!(
+                "AUDIO_USE_LEGACY_FEC active — reverting Opus encoder to Phase 0 \
+                 behavior (inband FEC Mode1, no DRED)"
+            );
+        }
 
         let mut enc = Self {
             inner: encoder,
             codec_id: profile.codec,
             frame_duration_ms: profile.frame_duration_ms,
+            legacy_fec_mode,
         };
-        enc.apply_bitrate(profile.codec)?;
-        enc.set_inband_fec(true);
-        enc.set_dtx(true);
 
-        // Voice signal type hint for better compression
+        // Common setup — bitrate, DTX, signal hint, complexity. These are
+        // identical regardless of the protection mode below.
+        enc.apply_bitrate(profile.codec)?;
+        enc.set_dtx(true);
         enc.inner
             .set_signal(Signal::Voice)
-            .map_err(|e| CodecError::EncodeFailed(format!("set signal: {e}")))?;
-
-        // Default complexity 7 — good quality/CPU trade-off for VoIP
+            .map_err(|e| CodecError::EncodeFailed(format!("set signal: {e:?}")))?;
         enc.inner
             .set_complexity(7)
-            .map_err(|e| CodecError::EncodeFailed(format!("set complexity: {e}")))?;
+            .map_err(|e| CodecError::EncodeFailed(format!("set complexity: {e:?}")))?;
+
+        // Protection mode: DRED (Phase 1 default) or legacy inband FEC.
+        enc.apply_protection_mode(profile.codec)?;
 
         Ok(enc)
     }
 
-    fn apply_bitrate(&mut self, codec: CodecId) -> Result<(), CodecError> {
-        let bps = codec.bitrate_bps() as i32;
+    /// Configure the protection mode for the active codec.
+    ///
+    /// In DRED mode (default): disable inband FEC, set DRED duration for the
+    /// codec tier, clamp packet_loss to the 5% floor so DRED stays active.
+    ///
+    /// In legacy mode: enable inband FEC Mode1 (Phase 0 behavior), leave
+    /// DRED and packet_loss at libopus defaults.
+    fn apply_protection_mode(&mut self, codec: CodecId) -> Result<(), CodecError> {
+        if self.legacy_fec_mode {
+            self.inner
+                .set_inband_fec(InbandFec::Mode1)
+                .map_err(|e| CodecError::EncodeFailed(format!("set inband FEC: {e:?}")))?;
+            // Leave DRED at 0 and packet_loss at default — matches Phase 0.
+            return Ok(());
+        }
+
+        // DRED path: disable the overlapping inband FEC, enable DRED with
+        // per-profile duration, floor packet_loss so DRED emits.
         self.inner
-            .set_bitrate(Bitrate::BitsPerSecond(bps))
-            .map_err(|e| CodecError::EncodeFailed(format!("set bitrate: {e}")))?;
+            .set_inband_fec(InbandFec::Off)
+            .map_err(|e| CodecError::EncodeFailed(format!("set inband FEC off: {e:?}")))?;
+
+        let dred_frames = dred_duration_for(codec);
+        self.inner
+            .set_dred_duration(dred_frames)
+            .map_err(|e| CodecError::EncodeFailed(format!("set DRED duration: {e:?}")))?;
+
+        self.inner
+            .set_packet_loss(DRED_LOSS_FLOOR_PCT)
+            .map_err(|e| CodecError::EncodeFailed(format!("set packet loss floor: {e:?}")))?;
+
+        // Both of these are gated behind the GUI debug toggle so logcat
+        // stays clean in normal mode. Flip "DRED verbose logs" in the
+        // settings panel to see the per-encoder config + libopus version.
+        if crate::dred_verbose_logs() {
+            info!(
+                codec = ?codec,
+                dred_frames,
+                dred_ms = dred_frames as u32 * 10,
+                loss_floor_pct = DRED_LOSS_FLOOR_PCT,
+                "opus encoder: DRED enabled"
+            );
+
+            // One-shot logging of the linked libopus version so we can
+            // confirm at a glance that opusic-c (libopus 1.5.2) is loaded.
+            // Pre-Phase-0 audiopus shipped libopus 1.3 which has no DRED;
+            // if this log says "libopus 1.3" something is very wrong.
+            LIBOPUS_VERSION_LOGGED.get_or_init(|| {
+                info!(libopus_version = %opusic_c::version(), "linked libopus version");
+            });
+        }
+
+        Ok(())
+    }
+
+    fn apply_bitrate(&mut self, codec: CodecId) -> Result<(), CodecError> {
+        let bps = codec.bitrate_bps();
+        self.inner
+            .set_bitrate(Bitrate::Value(bps))
+            .map_err(|e| CodecError::EncodeFailed(format!("set bitrate: {e:?}")))?;
         debug!(bitrate_bps = bps, "opus encoder bitrate set");
         Ok(())
     }
@@ -71,10 +238,36 @@ impl OpusEncoder {
 
     /// Hint the encoder about expected packet loss percentage (0-100).
     ///
-    /// Higher values cause the encoder to use more redundancy to survive
-    /// packet loss, at the expense of slightly higher bitrate.
+    /// In DRED mode, the value is floored at `DRED_LOSS_FLOOR_PCT` so the
+    /// encoder never drops DRED emission even on a perfect network. Real
+    /// loss measurements from the quality adapter override upward.
+    ///
+    /// In legacy mode, the value is passed through unchanged (min 0, max 100).
     pub fn set_expected_loss(&mut self, loss_pct: u8) {
-        let _ = self.inner.set_packet_loss_perc(loss_pct.min(100));
+        let clamped = if self.legacy_fec_mode {
+            loss_pct.min(100)
+        } else {
+            loss_pct.max(DRED_LOSS_FLOOR_PCT).min(100)
+        };
+        let _ = self.inner.set_packet_loss(clamped);
+    }
+
+    /// Set the DRED duration in 10 ms frame units (0 disables, max 104).
+    ///
+    /// No-op in legacy mode. Normally driven automatically by the active
+    /// quality profile via `apply_protection_mode`; this setter exists for
+    /// tests and for the rare case where a caller needs to override the
+    /// per-profile default.
+    pub fn set_dred_duration(&mut self, frames: u8) {
+        if self.legacy_fec_mode {
+            return;
+        }
+        let _ = self.inner.set_dred_duration(frames.min(104));
+    }
+
+    /// Test/introspection accessor: whether legacy FEC mode is active.
+    pub fn is_legacy_fec_mode(&self) -> bool {
+        self.legacy_fec_mode
     }
 }
 
@@ -87,10 +280,14 @@ impl AudioEncoder for OpusEncoder {
                 pcm.len()
             )));
         }
+        // opusic-c takes &[u16] for the sample input. Bit pattern is
+        // identical to i16 — the cast is zero-cost and the encoder
+        // interprets the bytes the same way as libopus internally.
+        let pcm_u16: &[u16] = bytemuck::cast_slice(pcm);
         let n = self
             .inner
-            .encode(pcm, out)
-            .map_err(|e| CodecError::EncodeFailed(format!("opus encode: {e}")))?;
+            .encode_to_slice(pcm_u16, out)
+            .map_err(|e| CodecError::EncodeFailed(format!("opus encode: {e:?}")))?;
         Ok(n)
     }
 
@@ -104,6 +301,9 @@ impl AudioEncoder for OpusEncoder {
                 self.codec_id = profile.codec;
                 self.frame_duration_ms = profile.frame_duration_ms;
                 self.apply_bitrate(profile.codec)?;
+                // Refresh DRED duration for the new tier. apply_protection_mode
+                // is idempotent and handles the legacy-vs-DRED branch correctly.
+                self.apply_protection_mode(profile.codec)?;
                 Ok(())
             }
             other => Err(CodecError::UnsupportedTransition {
@@ -120,10 +320,198 @@ impl AudioEncoder for OpusEncoder {
     }
 
     fn set_inband_fec(&mut self, enabled: bool) {
-        let _ = self.inner.set_inband_fec(enabled);
+        // In DRED mode, ignore external requests to re-enable inband FEC —
+        // running both mechanisms wastes bitrate on overlapping protection
+        // and opusic-c's own docs recommend disabling inband FEC when DRED
+        // is on. Trait callers that genuinely want classical FEC should set
+        // `AUDIO_USE_LEGACY_FEC=1` and re-create the encoder.
+        if !self.legacy_fec_mode {
+            debug!(
+                enabled,
+                "set_inband_fec ignored: DRED mode is active (set AUDIO_USE_LEGACY_FEC to revert)"
+            );
+            return;
+        }
+        let mode = if enabled { InbandFec::Mode1 } else { InbandFec::Off };
+        let _ = self.inner.set_inband_fec(mode);
     }
 
     fn set_dtx(&mut self, enabled: bool) {
         let _ = self.inner.set_dtx(enabled);
     }
+
+    fn set_expected_loss(&mut self, loss_pct: u8) {
+        OpusEncoder::set_expected_loss(self, loss_pct);
+    }
+
+    fn set_dred_duration(&mut self, frames: u8) {
+        OpusEncoder::set_dred_duration(self, frames);
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use wzp_proto::AudioDecoder;
+
+    /// Phase 0 acceptance gate: fail loudly if the linked libopus is not 1.5.x.
+    /// DRED (Phase 1+) only exists in libopus ≥ 1.5, so running against an
+    /// older version would silently regress the entire DRED integration.
+    #[test]
+    fn linked_libopus_is_1_5() {
+        let version = opusic_c::version();
+        assert!(
+            version.contains("1.5"),
+            "expected libopus 1.5.x, got: {version}"
+        );
+    }
+
+    #[test]
+    fn encoder_creates_at_good_profile() {
+        let enc = OpusEncoder::new(QualityProfile::GOOD).expect("opus encoder init");
+        assert_eq!(enc.codec_id, CodecId::Opus24k);
+        assert_eq!(enc.frame_samples(), 960); // 20 ms @ 48 kHz
+    }
+
+    #[test]
+    fn encoder_roundtrip_silence() {
+        let mut enc = OpusEncoder::new(QualityProfile::GOOD).unwrap();
+        let mut dec = crate::opus_dec::OpusDecoder::new(QualityProfile::GOOD).unwrap();
+        let pcm_in = vec![0i16; 960]; // 20 ms silence
+        let mut encoded = vec![0u8; 512];
+        let n = enc.encode(&pcm_in, &mut encoded).unwrap();
+        assert!(n > 0);
+        let mut pcm_out = vec![0i16; 960];
+        let samples = dec.decode(&encoded[..n], &mut pcm_out).unwrap();
+        assert_eq!(samples, 960);
+    }
+
+    // ─── Phase 1 — DRED duration policy ─────────────────────────────────────
+
+    #[test]
+    fn dred_duration_for_studio_tiers_is_100ms() {
+        assert_eq!(dred_duration_for(CodecId::Opus32k), 10);
+        assert_eq!(dred_duration_for(CodecId::Opus48k), 10);
+        assert_eq!(dred_duration_for(CodecId::Opus64k), 10);
+    }
+
+    #[test]
+    fn dred_duration_for_normal_tiers_is_200ms() {
+        assert_eq!(dred_duration_for(CodecId::Opus16k), 20);
+        assert_eq!(dred_duration_for(CodecId::Opus24k), 20);
+    }
+
+    #[test]
+    fn dred_duration_for_degraded_tier_is_1040ms() {
+        assert_eq!(dred_duration_for(CodecId::Opus6k), 104);
+    }
+
+    #[test]
+    fn dred_duration_for_codec2_is_zero() {
+        assert_eq!(dred_duration_for(CodecId::Codec2_3200), 0);
+        assert_eq!(dred_duration_for(CodecId::Codec2_1200), 0);
+        assert_eq!(dred_duration_for(CodecId::ComfortNoise), 0);
+    }
+
+    // ─── Phase 1 — Legacy escape hatch ──────────────────────────────────────
+
+    /// By default (env var unset), legacy mode is off.
+    ///
+    /// This test does NOT manipulate the environment to avoid flakiness
+    /// when the full suite runs in parallel. It only asserts on a freshly
+    /// created encoder in the ambient environment.
+    #[test]
+    fn default_mode_is_dred_not_legacy() {
+        // SAFETY: only run if the ambient env hasn't set the var externally.
+        if std::env::var(LEGACY_FEC_ENV).is_ok() {
+            return; // don't assert — someone set the env for a reason.
+        }
+        let enc = OpusEncoder::new(QualityProfile::GOOD).unwrap();
+        assert!(!enc.is_legacy_fec_mode());
+    }
+
+    // ─── Phase 1 — Behavioral regression: roundtrip still works ─────────────
+
+    #[test]
+    fn dred_mode_roundtrip_voice_pattern() {
+        // Use a realistic voice-like input (sine wave at speech frequencies)
+        // so the encoder emits meaningful DRED data rather than trivially
+        // compressible silence.
+        let mut enc = OpusEncoder::new(QualityProfile::GOOD).unwrap();
+        let mut dec = crate::opus_dec::OpusDecoder::new(QualityProfile::GOOD).unwrap();
+
+        let mut total_encoded_bytes = 0usize;
+        // Run 50 frames (1 second) so DRED fills up and starts emitting.
+        for frame_idx in 0..50 {
+            let pcm_in: Vec<i16> = (0..960)
+                .map(|i| {
+                    let t = (frame_idx * 960 + i) as f64 / 48_000.0;
+                    (8000.0 * (2.0 * std::f64::consts::PI * 300.0 * t).sin()) as i16
+                })
+                .collect();
+            let mut encoded = vec![0u8; 512];
+            let n = enc.encode(&pcm_in, &mut encoded).unwrap();
+            assert!(n > 0);
+            total_encoded_bytes += n;
+
+            let mut pcm_out = vec![0i16; 960];
+            let samples = dec.decode(&encoded[..n], &mut pcm_out).unwrap();
+            assert_eq!(samples, 960);
+        }
+
+        // Effective bitrate after 1 second of encoding.
+        // Opus 24k base + ~1 kbps DRED ≈ 25 kbps ≈ 3125 bytes/sec.
+        // Allow generous headroom (2000 lower bound, 8000 upper bound) —
+        // this is a behavioral regression check, not a tight bitrate assertion.
+        // The exact value is printed with --nocapture for diagnostic use.
+        eprintln!(
+            "[phase1 bitrate probe] legacy_fec_mode={} total_encoded={} bytes/sec",
+            enc.is_legacy_fec_mode(),
+            total_encoded_bytes
+        );
+        assert!(
+            total_encoded_bytes > 2000,
+            "encoder output too small: {total_encoded_bytes} bytes/sec (DRED likely not emitting)"
+        );
+        assert!(
+            total_encoded_bytes < 8000,
+            "encoder output too large: {total_encoded_bytes} bytes/sec"
+        );
+    }
+
+    // ─── Phase 1 — set_profile updates DRED duration on tier switch ─────────
+
+    #[test]
+    fn profile_switch_refreshes_dred_duration() {
+        // Start on GOOD (Opus 24k, DRED 20 frames), switch to DEGRADED
+        // (Opus 6k, DRED 50 frames). The encoder should accept both profile
+        // changes without error. We can't directly observe the DRED duration
+        // inside libopus, but apply_protection_mode returns Ok for both.
+        let mut enc = OpusEncoder::new(QualityProfile::GOOD).unwrap();
+        assert_eq!(enc.codec_id, CodecId::Opus24k);
+
+        enc.set_profile(QualityProfile::DEGRADED).unwrap();
+        assert_eq!(enc.codec_id, CodecId::Opus6k);
+
+        enc.set_profile(QualityProfile::STUDIO_64K).unwrap();
+        assert_eq!(enc.codec_id, CodecId::Opus64k);
+    }
+
+    // ─── Phase 1 — Trait set_inband_fec is a no-op in DRED mode ─────────────
+
+    #[test]
+    fn set_inband_fec_noop_in_dred_mode() {
+        if std::env::var(LEGACY_FEC_ENV).is_ok() {
+            return;
+        }
+        let mut enc = OpusEncoder::new(QualityProfile::GOOD).unwrap();
+        // Should not error, should not re-enable inband FEC internally.
+        enc.set_inband_fec(true);
+        // We can't directly query libopus's inband FEC state through opusic-c,
+        // but the call must not panic and the encoder must still work.
+        let pcm_in = vec![0i16; 960];
+        let mut encoded = vec![0u8; 512];
+        let n = enc.encode(&pcm_in, &mut encoded).unwrap();
+        assert!(n > 0);
+    }
 }

crates/wzp-crypto/src/handshake.rs

@@ -18,10 +18,14 @@ use crate::session::ChaChaSession;
 pub struct WarzoneKeyExchange {
     /// Ed25519 signing key (identity).
     signing_key: SigningKey,
-    /// X25519 static secret (derived from seed, used for identity encryption).
+    /// X25519 static secret derived from identity seed. Reserved for future
+    /// use in static-key federation authentication (not used in current
+    /// ephemeral-only handshake protocol).
     #[allow(dead_code)]
     x25519_static_secret: StaticSecret,
-    /// X25519 static public key.
+    /// X25519 static public key derived from identity seed. Reserved for
+    /// future use in static-key federation authentication (not used in
+    /// current ephemeral-only handshake protocol).
     #[allow(dead_code)]
     x25519_static_public: X25519PublicKey,
     /// Ephemeral X25519 secret for the current call (set by generate_ephemeral).

crates/wzp-crypto/tests/featherchat_compat.rs

@@ -115,6 +115,7 @@ fn wzp_signal_serializes_into_fc_callsignal_payload() {
         ephemeral_pub: [2u8; 32],
         signature: vec![3u8; 64],
         supported_profiles: vec![wzp_proto::QualityProfile::GOOD],
+        alias: None,
     };
 
     // Encode as featherChat CallSignal payload
@@ -198,6 +199,7 @@ fn wzp_answer_round_trips_through_fc_callsignal() {
 fn wzp_hangup_round_trips_through_fc_callsignal() {
     let hangup = wzp_proto::SignalMessage::Hangup {
         reason: wzp_proto::HangupReason::Normal,
+        call_id: None,
     };
 
     let payload = wzp_client::featherchat::encode_call_payload(&hangup, None, None);
@@ -273,13 +275,14 @@ fn auth_invalid_response_matches() {
 
 #[test]
 fn all_signal_types_map_correctly() {
-    use wzp_client::featherchat::{signal_to_call_type, CallSignalType};
+    use wzp_client::featherchat::signal_to_call_type;
 
     let cases: Vec<(wzp_proto::SignalMessage, &str)> = vec![
         (
             wzp_proto::SignalMessage::CallOffer {
                 identity_pub: [0; 32], ephemeral_pub: [0; 32],
                 signature: vec![], supported_profiles: vec![],
+                alias: None,
             },
             "Offer",
         ),
@@ -300,6 +303,7 @@ fn all_signal_types_map_correctly() {
         (
             wzp_proto::SignalMessage::Hangup {
                 reason: wzp_proto::HangupReason::Normal,
+                call_id: None,
             },
             "Hangup",
         ),

crates/wzp-native/Cargo.toml

@@ -0,0 +1,29 @@
+[package]
+name = "wzp-native"
+version = "0.1.0"
+edition = "2024"
+description = "WarzonePhone native audio library — standalone Android cdylib that eventually owns all C++ (Oboe bridge) and exposes a pure-C FFI. Built with cargo-ndk, loaded at runtime by the Tauri desktop cdylib via libloading."
+
+# Crate-type is DELIBERATELY only cdylib (no rlib, no staticlib). This crate
+# is built with `cargo ndk -t arm64-v8a build --release -p wzp-native` as a
+# standalone .so, which is the same path the legacy wzp-android crate uses
+# successfully on the same phone / same NDK. Keeping the crate-type single
+# avoids the rust-lang/rust#104707 symbol leak that bit us when Tauri's
+# desktop crate had ["staticlib", "cdylib", "rlib"] and any C++ static
+# archive pulled bionic's internal pthread_create into the final .so.
+[lib]
+name = "wzp_native"
+crate-type = ["cdylib"]
+
+[build-dependencies]
+# cc is SAFE to use here because this crate is a single-cdylib: no
+# staticlib in crate-type → no rust-lang/rust#104707 symbol leak. The
+# legacy wzp-android crate uses the same setup and works.
+cc = "1"
+
+[dependencies]
+# Phase 2: Oboe C++ audio bridge. Still no Rust deps — we do the whole
+# audio pipeline via extern "C" into the bundled C++ and expose our own
+# narrow extern "C" API for wzp-desktop to dlopen via libloading.
+# Phase 3 can add wzp-proto/wzp-codec if we want to share codec logic
+# instead of calling back into wzp-desktop via callbacks.

crates/wzp-native/build.rs

@@ -0,0 +1,119 @@
+//! wzp-native build.rs — Oboe C++ bridge compile on Android.
+//!
+//! Near-verbatim copy of crates/wzp-android/build.rs (which is known to
+//! work). The crucial distinction: this crate is a single-cdylib (no
+//! staticlib, no rlib in crate-type) so rust-lang/rust#104707 doesn't
+//! apply — bionic's internal pthread_create / __init_tcb symbols stay
+//! UND and resolve against libc.so at runtime, as they should.
+//!
+//! On non-Android hosts we compile `cpp/oboe_stub.cpp` (empty stubs) so
+//! `cargo check --target <host>` still works for IDEs and CI.
+
+use std::path::PathBuf;
+
+fn main() {
+    let target = std::env::var("TARGET").unwrap_or_default();
+
+    if target.contains("android") {
+        // getauxval_fix: override compiler-rt's broken static getauxval
+        // stub that SIGSEGVs in shared libraries.
+        cc::Build::new()
+            .file("cpp/getauxval_fix.c")
+            .compile("wzp_native_getauxval_fix");
+
+        let oboe_dir = fetch_oboe();
+        match oboe_dir {
+            Some(oboe_path) => {
+                println!("cargo:warning=wzp-native: building with Oboe from {:?}", oboe_path);
+                let mut build = cc::Build::new();
+                build
+                    .cpp(true)
+                    .std("c++17")
+                    // Shared libc++ — matches legacy wzp-android setup.
+                    .cpp_link_stdlib(Some("c++_shared"))
+                    .include("cpp")
+                    .include(oboe_path.join("include"))
+                    .include(oboe_path.join("src"))
+                    .define("WZP_HAS_OBOE", None)
+                    .file("cpp/oboe_bridge.cpp");
+                add_cpp_files_recursive(&mut build, &oboe_path.join("src"));
+                build.compile("wzp_native_oboe_bridge");
+            }
+            None => {
+                println!("cargo:warning=wzp-native: Oboe not found, building stub");
+                cc::Build::new()
+                    .cpp(true)
+                    .std("c++17")
+                    .cpp_link_stdlib(Some("c++_shared"))
+                    .file("cpp/oboe_stub.cpp")
+                    .include("cpp")
+                    .compile("wzp_native_oboe_bridge");
+            }
+        }
+
+        // Oboe needs log + OpenSLES backends at runtime.
+        println!("cargo:rustc-link-lib=log");
+        println!("cargo:rustc-link-lib=OpenSLES");
+
+        // Re-run if any cpp file changes
+        println!("cargo:rerun-if-changed=cpp/oboe_bridge.cpp");
+        println!("cargo:rerun-if-changed=cpp/oboe_bridge.h");
+        println!("cargo:rerun-if-changed=cpp/oboe_stub.cpp");
+        println!("cargo:rerun-if-changed=cpp/getauxval_fix.c");
+    } else {
+        // Non-Android hosts: compile the empty stub so lib.rs's extern
+        // declarations resolve when someone runs `cargo check` on macOS
+        // or Linux without an NDK.
+        cc::Build::new()
+            .cpp(true)
+            .std("c++17")
+            .file("cpp/oboe_stub.cpp")
+            .include("cpp")
+            .compile("wzp_native_oboe_bridge");
+        println!("cargo:rerun-if-changed=cpp/oboe_stub.cpp");
+    }
+}
+
+/// Recursively add all `.cpp` files from a directory to a cc::Build.
+fn add_cpp_files_recursive(build: &mut cc::Build, dir: &std::path::Path) {
+    if !dir.is_dir() {
+        return;
+    }
+    for entry in std::fs::read_dir(dir).unwrap() {
+        let entry = entry.unwrap();
+        let path = entry.path();
+        if path.is_dir() {
+            add_cpp_files_recursive(build, &path);
+        } else if path.extension().map_or(false, |e| e == "cpp") {
+            build.file(&path);
+        }
+    }
+}
+
+/// Fetch or find Oboe headers + sources (v1.8.1). Same logic as the
+/// legacy wzp-android crate's build.rs.
+fn fetch_oboe() -> Option<PathBuf> {
+    let out_dir = PathBuf::from(std::env::var("OUT_DIR").unwrap());
+    let oboe_dir = out_dir.join("oboe");
+
+    if oboe_dir.join("include").join("oboe").join("Oboe.h").exists() {
+        return Some(oboe_dir);
+    }
+
+    let status = std::process::Command::new("git")
+        .args([
+            "clone",
+            "--depth=1",
+            "--branch=1.8.1",
+            "https://github.com/google/oboe.git",
+            oboe_dir.to_str().unwrap(),
+        ])
+        .status();
+
+    match status {
+        Ok(s) if s.success() && oboe_dir.join("include").join("oboe").join("Oboe.h").exists() => {
+            Some(oboe_dir)
+        }
+        _ => None,
+    }
+}

crates/wzp-native/cpp/getauxval_fix.c

@@ -0,0 +1,21 @@
+// Override the broken static getauxval from compiler-rt/CRT.
+// The static version reads from __libc_auxv which is NULL in shared libs
+// loaded via dlopen, causing SIGSEGV in init_have_lse_atomics at load time.
+// This version calls the real bionic getauxval via dlsym.
+#ifdef __ANDROID__
+#include <dlfcn.h>
+#include <stdint.h>
+
+typedef unsigned long (*getauxval_fn)(unsigned long);
+
+unsigned long getauxval(unsigned long type) {
+    static getauxval_fn real_getauxval = (getauxval_fn)0;
+    if (!real_getauxval) {
+        real_getauxval = (getauxval_fn)dlsym((void*)-1L /* RTLD_DEFAULT */, "getauxval");
+        if (!real_getauxval) {
+            return 0;
+        }
+    }
+    return real_getauxval(type);
+}
+#endif

crates/wzp-native/cpp/oboe_bridge.cpp

@@ -0,0 +1,477 @@
+// Full Oboe implementation for Android
+// This file is compiled only when targeting Android
+
+#include "oboe_bridge.h"
+
+#ifdef __ANDROID__
+#include <oboe/Oboe.h>
+#include <android/log.h>
+#include <cstring>
+#include <atomic>
+#include <chrono>
+#include <thread>
+
+#define LOG_TAG "wzp-oboe"
+#define LOGI(...) __android_log_print(ANDROID_LOG_INFO, LOG_TAG, __VA_ARGS__)
+#define LOGW(...) __android_log_print(ANDROID_LOG_WARN, LOG_TAG, __VA_ARGS__)
+#define LOGE(...) __android_log_print(ANDROID_LOG_ERROR, LOG_TAG, __VA_ARGS__)
+
+// ---------------------------------------------------------------------------
+// Ring buffer helpers (SPSC, lock-free)
+// ---------------------------------------------------------------------------
+
+static inline int32_t ring_available_read(const wzp_atomic_int* write_idx,
+                                           const wzp_atomic_int* read_idx,
+                                           int32_t capacity) {
+    int32_t w = std::atomic_load_explicit(write_idx, std::memory_order_acquire);
+    int32_t r = std::atomic_load_explicit(read_idx, std::memory_order_relaxed);
+    int32_t avail = w - r;
+    if (avail < 0) avail += capacity;
+    return avail;
+}
+
+static inline int32_t ring_available_write(const wzp_atomic_int* write_idx,
+                                            const wzp_atomic_int* read_idx,
+                                            int32_t capacity) {
+    return capacity - 1 - ring_available_read(write_idx, read_idx, capacity);
+}
+
+static inline void ring_write(int16_t* buf, int32_t capacity,
+                               wzp_atomic_int* write_idx, const wzp_atomic_int* read_idx,
+                               const int16_t* src, int32_t count) {
+    int32_t w = std::atomic_load_explicit(write_idx, std::memory_order_relaxed);
+    for (int32_t i = 0; i < count; i++) {
+        buf[w] = src[i];
+        w++;
+        if (w >= capacity) w = 0;
+    }
+    std::atomic_store_explicit(write_idx, w, std::memory_order_release);
+}
+
+static inline void ring_read(int16_t* buf, int32_t capacity,
+                              const wzp_atomic_int* write_idx, wzp_atomic_int* read_idx,
+                              int16_t* dst, int32_t count) {
+    int32_t r = std::atomic_load_explicit(read_idx, std::memory_order_relaxed);
+    for (int32_t i = 0; i < count; i++) {
+        dst[i] = buf[r];
+        r++;
+        if (r >= capacity) r = 0;
+    }
+    std::atomic_store_explicit(read_idx, r, std::memory_order_release);
+}
+
+// ---------------------------------------------------------------------------
+// Global state
+// ---------------------------------------------------------------------------
+
+static std::shared_ptr<oboe::AudioStream> g_capture_stream;
+static std::shared_ptr<oboe::AudioStream> g_playout_stream;
+// Value copy — the WzpOboeRings the Rust side passes us lives on the caller's
+// stack frame and goes away as soon as wzp_oboe_start returns. The raw
+// int16/atomic pointers INSIDE the struct point into the Rust-owned, leaked-
+// for-the-lifetime-of-the-process AudioBackend singleton, so copying the
+// struct by value is safe and keeps the inner pointers valid indefinitely.
+// g_rings_valid guards the audio-callback-side read; clearing it in stop()
+// signals "no backend" to the callbacks which then return silence + Stop.
+static WzpOboeRings g_rings{};
+static std::atomic<bool> g_rings_valid{false};
+static std::atomic<bool> g_running{false};
+static std::atomic<float> g_capture_latency_ms{0.0f};
+static std::atomic<float> g_playout_latency_ms{0.0f};
+
+// ---------------------------------------------------------------------------
+// Capture callback
+// ---------------------------------------------------------------------------
+
+class CaptureCallback : public oboe::AudioStreamDataCallback {
+public:
+    uint64_t calls = 0;
+    uint64_t total_frames = 0;
+    uint64_t total_written = 0;
+    uint64_t ring_full_drops = 0;
+
+    oboe::DataCallbackResult onAudioReady(
+            oboe::AudioStream* stream,
+            void* audioData,
+            int32_t numFrames) override {
+        if (!g_running.load(std::memory_order_relaxed) ||
+            !g_rings_valid.load(std::memory_order_acquire)) {
+            return oboe::DataCallbackResult::Stop;
+        }
+
+        const int16_t* src = static_cast<const int16_t*>(audioData);
+        int32_t avail = ring_available_write(g_rings.capture_write_idx,
+                                              g_rings.capture_read_idx,
+                                              g_rings.capture_capacity);
+        int32_t to_write = (numFrames < avail) ? numFrames : avail;
+        if (to_write > 0) {
+            ring_write(g_rings.capture_buf, g_rings.capture_capacity,
+                       g_rings.capture_write_idx, g_rings.capture_read_idx,
+                       src, to_write);
+        }
+        total_frames += numFrames;
+        total_written += to_write;
+        if (to_write < numFrames) {
+            ring_full_drops += (numFrames - to_write);
+        }
+
+        // Sample-range probe on the FIRST callback to prove we get real audio
+        if (calls == 0 && numFrames > 0) {
+            int16_t lo = src[0], hi = src[0];
+            int32_t sumsq = 0;
+            for (int32_t i = 0; i < numFrames; i++) {
+                if (src[i] < lo) lo = src[i];
+                if (src[i] > hi) hi = src[i];
+                sumsq += (int32_t)src[i] * (int32_t)src[i];
+            }
+            int32_t rms = (int32_t) (numFrames > 0 ? (int32_t)__builtin_sqrt((double)sumsq / (double)numFrames) : 0);
+            LOGI("capture cb#0: numFrames=%d sample_range=[%d..%d] rms=%d to_write=%d",
+                 numFrames, lo, hi, rms, to_write);
+        }
+        // Heartbeat every 50 callbacks (~1s at 20ms/burst)
+        calls++;
+        if ((calls % 50) == 0) {
+            LOGI("capture heartbeat: calls=%llu numFrames=%d ring_avail_write=%d to_write=%d full_drops=%llu total_written=%llu",
+                 (unsigned long long)calls, numFrames, avail, to_write,
+                 (unsigned long long)ring_full_drops, (unsigned long long)total_written);
+        }
+
+        // Update latency estimate
+        auto result = stream->calculateLatencyMillis();
+        if (result) {
+            g_capture_latency_ms.store(static_cast<float>(result.value()),
+                                        std::memory_order_relaxed);
+        }
+
+        return oboe::DataCallbackResult::Continue;
+    }
+};
+
+// ---------------------------------------------------------------------------
+// Playout callback
+// ---------------------------------------------------------------------------
+
+class PlayoutCallback : public oboe::AudioStreamDataCallback {
+public:
+    uint64_t calls = 0;
+    uint64_t total_frames = 0;
+    uint64_t total_played_real = 0;
+    uint64_t underrun_frames = 0;
+    uint64_t nonempty_calls = 0;
+
+    oboe::DataCallbackResult onAudioReady(
+            oboe::AudioStream* stream,
+            void* audioData,
+            int32_t numFrames) override {
+        if (!g_running.load(std::memory_order_relaxed) ||
+            !g_rings_valid.load(std::memory_order_acquire)) {
+            memset(audioData, 0, numFrames * sizeof(int16_t));
+            return oboe::DataCallbackResult::Stop;
+        }
+
+        int16_t* dst = static_cast<int16_t*>(audioData);
+        int32_t avail = ring_available_read(g_rings.playout_write_idx,
+                                             g_rings.playout_read_idx,
+                                             g_rings.playout_capacity);
+        int32_t to_read = (numFrames < avail) ? numFrames : avail;
+
+        if (to_read > 0) {
+            ring_read(g_rings.playout_buf, g_rings.playout_capacity,
+                      g_rings.playout_write_idx, g_rings.playout_read_idx,
+                      dst, to_read);
+            nonempty_calls++;
+        }
+        // Fill remainder with silence on underrun
+        if (to_read < numFrames) {
+            memset(dst + to_read, 0, (numFrames - to_read) * sizeof(int16_t));
+            underrun_frames += (numFrames - to_read);
+        }
+        total_frames += numFrames;
+        total_played_real += to_read;
+
+        // First callback: log requested config + prove we're being called
+        if (calls == 0) {
+            LOGI("playout cb#0: numFrames=%d ring_avail_read=%d to_read=%d",
+                 numFrames, avail, to_read);
+        }
+        // On the first callback that actually has data, log the sample range
+        // so we can tell if the samples coming out of the ring look like real
+        // audio vs constant-zeroes vs garbage.
+        if (to_read > 0 && nonempty_calls == 1) {
+            int16_t lo = dst[0], hi = dst[0];
+            int32_t sumsq = 0;
+            for (int32_t i = 0; i < to_read; i++) {
+                if (dst[i] < lo) lo = dst[i];
+                if (dst[i] > hi) hi = dst[i];
+                sumsq += (int32_t)dst[i] * (int32_t)dst[i];
+            }
+            int32_t rms = (to_read > 0) ? (int32_t)__builtin_sqrt((double)sumsq / (double)to_read) : 0;
+            LOGI("playout FIRST nonempty read: to_read=%d sample_range=[%d..%d] rms=%d",
+                 to_read, lo, hi, rms);
+        }
+        // Heartbeat every 50 callbacks (~1s at 20ms/burst)
+        calls++;
+        if ((calls % 50) == 0) {
+            int state = (int)stream->getState();
+            auto xrunRes = stream->getXRunCount();
+            int xruns = xrunRes ? xrunRes.value() : -1;
+            LOGI("playout heartbeat: calls=%llu nonempty=%llu numFrames=%d ring_avail_read=%d to_read=%d underrun_frames=%llu total_played_real=%llu state=%d xruns=%d",
+                 (unsigned long long)calls, (unsigned long long)nonempty_calls,
+                 numFrames, avail, to_read,
+                 (unsigned long long)underrun_frames, (unsigned long long)total_played_real,
+                 state, xruns);
+        }
+
+        // Update latency estimate
+        auto result = stream->calculateLatencyMillis();
+        if (result) {
+            g_playout_latency_ms.store(static_cast<float>(result.value()),
+                                        std::memory_order_relaxed);
+        }
+
+        return oboe::DataCallbackResult::Continue;
+    }
+};
+
+static CaptureCallback g_capture_cb;
+static PlayoutCallback g_playout_cb;
+
+// ---------------------------------------------------------------------------
+// Public C API
+// ---------------------------------------------------------------------------
+
+int wzp_oboe_start(const WzpOboeConfig* config, const WzpOboeRings* rings) {
+    if (g_running.load(std::memory_order_relaxed)) {
+        LOGW("wzp_oboe_start: already running");
+        return -1;
+    }
+
+    // Deep-copy the rings struct into static storage BEFORE we publish it to
+    // the audio callbacks — `rings` points at the caller's stack frame and
+    // goes away as soon as this function returns.
+    g_rings = *rings;
+    g_rings_valid.store(true, std::memory_order_release);
+
+    // Build capture stream
+    oboe::AudioStreamBuilder captureBuilder;
+    captureBuilder.setDirection(oboe::Direction::Input)
+        ->setPerformanceMode(oboe::PerformanceMode::LowLatency)
+        ->setSharingMode(oboe::SharingMode::Shared)
+        ->setFormat(oboe::AudioFormat::I16)
+        ->setChannelCount(config->channel_count)
+        ->setSampleRateConversionQuality(oboe::SampleRateConversionQuality::Best)
+        ->setDataCallback(&g_capture_cb);
+
+    if (config->bt_active) {
+        // BT SCO mode: do NOT set sample rate or input preset.
+        // Requesting 48kHz against a BT SCO device fails with
+        // "getInputProfile could not find profile". Letting the system
+        // choose the native rate (8/16kHz) and relying on Oboe's
+        // resampler (SampleRateConversionQuality::Best) to bridge
+        // to our 48kHz ring buffer is the only path that works.
+        // InputPreset::VoiceCommunication can also prevent BT SCO
+        // routing on some devices — skip it for BT.
+        LOGI("capture: BT mode — no sample rate or input preset set");
+    } else {
+        captureBuilder.setSampleRate(config->sample_rate)
+            ->setFramesPerDataCallback(config->frames_per_burst)
+            ->setInputPreset(oboe::InputPreset::VoiceCommunication);
+    }
+
+    oboe::Result result = captureBuilder.openStream(g_capture_stream);
+    if (result != oboe::Result::OK) {
+        LOGE("Failed to open capture stream: %s", oboe::convertToText(result));
+        return -2;
+    }
+    LOGI("capture stream opened: actualSR=%d actualCh=%d actualFormat=%d actualFramesPerBurst=%d actualFramesPerDataCallback=%d bufferCapacityInFrames=%d sharing=%d perfMode=%d",
+         g_capture_stream->getSampleRate(),
+         g_capture_stream->getChannelCount(),
+         (int)g_capture_stream->getFormat(),
+         g_capture_stream->getFramesPerBurst(),
+         g_capture_stream->getFramesPerDataCallback(),
+         g_capture_stream->getBufferCapacityInFrames(),
+         (int)g_capture_stream->getSharingMode(),
+         (int)g_capture_stream->getPerformanceMode());
+
+    // Build playout stream.
+    //
+    // Regression triangulation between builds:
+    //   96be740 (Usage::Media, default API): playout callback DID drain
+    //   the ring at steady 50Hz (playout heartbeat: calls=1100,
+    //   total_played_real=1055040). Audio not audible because OS routing
+    //   sent it to a silent output.
+    //
+    //   8c36fb5 (Usage::VoiceCommunication + setAudioApi(AAudio) +
+    //   ContentType::Speech): playout callback fired cb#0 once then
+    //   stopped draining the ring entirely. written_samples stuck at
+    //   ring capacity (7679) across all subsequent heartbeats, so Oboe
+    //   accepted zero samples after startup. Still inaudible.
+    //
+    // Hypothesis: forcing setAudioApi(AAudio) + VoiceCommunication on
+    // Pixel 6 / Android 15 opens a stream that succeeds at cb#0 but
+    // then detaches from the real audio driver. Reverting to the
+    // config that at least drove callbacks correctly, plus the
+    // Kotlin-side MODE_IN_COMMUNICATION + setSpeakerphoneOn(true)
+    // handled in MainActivity.kt to route audio to the loud speaker.
+    // Usage::VoiceCommunication is the correct Oboe usage for a VoIP app
+    // — it respects Android's in-call audio routing and lets
+    // AudioManager.setSpeakerphoneOn/setBluetoothScoOn actually switch
+    // between earpiece, loudspeaker, and Bluetooth headset. Combined with
+    // MODE_IN_COMMUNICATION set from MainActivity.kt and
+    // speakerphoneOn=false by default, this produces handset/earpiece as
+    // the default output.
+    //
+    // IMPORTANT: do NOT add setAudioApi(AAudio) here. Build 8c36fb5 proved
+    // forcing AAudio with Usage::VoiceCommunication makes the playout
+    // callback stop draining the ring after cb#0, even though the stream
+    // opens successfully. Letting Oboe pick the API (which will be AAudio
+    // on API ≥ 27 but via a different codepath) kept callbacks firing in
+    // every other build.
+    oboe::AudioStreamBuilder playoutBuilder;
+    playoutBuilder.setDirection(oboe::Direction::Output)
+        ->setPerformanceMode(oboe::PerformanceMode::LowLatency)
+        ->setSharingMode(oboe::SharingMode::Shared)
+        ->setFormat(oboe::AudioFormat::I16)
+        ->setChannelCount(config->channel_count)
+        ->setSampleRateConversionQuality(oboe::SampleRateConversionQuality::Best)
+        ->setDataCallback(&g_playout_cb);
+
+    if (config->bt_active) {
+        LOGI("playout: BT mode — no sample rate set, using Usage::Media");
+        // Usage::Media instead of VoiceCommunication for BT output
+        // to avoid conflicts with the communication device routing.
+        playoutBuilder.setUsage(oboe::Usage::Media);
+    } else {
+        playoutBuilder.setSampleRate(config->sample_rate)
+            ->setFramesPerDataCallback(config->frames_per_burst)
+            ->setUsage(oboe::Usage::VoiceCommunication);
+    }
+
+    result = playoutBuilder.openStream(g_playout_stream);
+    if (result != oboe::Result::OK) {
+        LOGE("Failed to open playout stream: %s", oboe::convertToText(result));
+        g_capture_stream->close();
+        g_capture_stream.reset();
+        return -3;
+    }
+    LOGI("playout stream opened: actualSR=%d actualCh=%d actualFormat=%d actualFramesPerBurst=%d actualFramesPerDataCallback=%d bufferCapacityInFrames=%d sharing=%d perfMode=%d",
+         g_playout_stream->getSampleRate(),
+         g_playout_stream->getChannelCount(),
+         (int)g_playout_stream->getFormat(),
+         g_playout_stream->getFramesPerBurst(),
+         g_playout_stream->getFramesPerDataCallback(),
+         g_playout_stream->getBufferCapacityInFrames(),
+         (int)g_playout_stream->getSharingMode(),
+         (int)g_playout_stream->getPerformanceMode());
+
+    g_running.store(true, std::memory_order_release);
+
+    // Start both streams
+    result = g_capture_stream->requestStart();
+    if (result != oboe::Result::OK) {
+        LOGE("Failed to start capture: %s", oboe::convertToText(result));
+        g_running.store(false, std::memory_order_release);
+        g_capture_stream->close();
+        g_playout_stream->close();
+        g_capture_stream.reset();
+        g_playout_stream.reset();
+        return -4;
+    }
+
+    result = g_playout_stream->requestStart();
+    if (result != oboe::Result::OK) {
+        LOGE("Failed to start playout: %s", oboe::convertToText(result));
+        g_running.store(false, std::memory_order_release);
+        g_capture_stream->requestStop();
+        g_capture_stream->close();
+        g_playout_stream->close();
+        g_capture_stream.reset();
+        g_playout_stream.reset();
+        return -5;
+    }
+
+    // Log initial stream states right after requestStart() returns.
+    // On well-behaved HALs both will already be Started; on others
+    // (Nothing A059) they may still be in Starting state.
+    LOGI("requestStart returned: capture_state=%d playout_state=%d",
+         (int)g_capture_stream->getState(),
+         (int)g_playout_stream->getState());
+
+    // Poll until both streams report Started state, up to 2s timeout.
+    // Some Android HALs (Nothing A059) delay transitioning from Starting
+    // to Started; proceeding before the transition completes causes the
+    // first capture/playout callbacks to be dropped silently.
+    {
+        auto deadline = std::chrono::steady_clock::now() + std::chrono::milliseconds(2000);
+        int poll_count = 0;
+        while (std::chrono::steady_clock::now() < deadline) {
+            auto cap_state = g_capture_stream->getState();
+            auto play_state = g_playout_stream->getState();
+            if (cap_state == oboe::StreamState::Started &&
+                play_state == oboe::StreamState::Started) {
+                LOGI("both streams Started after %d polls", poll_count);
+                break;
+            }
+            poll_count++;
+            std::this_thread::sleep_for(std::chrono::milliseconds(10));
+        }
+        // Log final state even on timeout (helps diagnose HAL quirks)
+        LOGI("stream states after poll: capture=%d playout=%d (polls=%d)",
+             (int)g_capture_stream->getState(),
+             (int)g_playout_stream->getState(),
+             poll_count);
+    }
+
+    LOGI("Oboe started: sr=%d burst=%d ch=%d",
+         config->sample_rate, config->frames_per_burst, config->channel_count);
+    return 0;
+}
+
+void wzp_oboe_stop(void) {
+    g_running.store(false, std::memory_order_release);
+    // Tell the audio callbacks to stop touching g_rings BEFORE we tear down
+    // the streams, so any in-flight callback returns Stop instead of reading
+    // stale pointers.
+    g_rings_valid.store(false, std::memory_order_release);
+
+    if (g_capture_stream) {
+        g_capture_stream->requestStop();
+        g_capture_stream->close();
+        g_capture_stream.reset();
+    }
+    if (g_playout_stream) {
+        g_playout_stream->requestStop();
+        g_playout_stream->close();
+        g_playout_stream.reset();
+    }
+
+    LOGI("Oboe stopped");
+}
+
+float wzp_oboe_capture_latency_ms(void) {
+    return g_capture_latency_ms.load(std::memory_order_relaxed);
+}
+
+float wzp_oboe_playout_latency_ms(void) {
+    return g_playout_latency_ms.load(std::memory_order_relaxed);
+}
+
+int wzp_oboe_is_running(void) {
+    return g_running.load(std::memory_order_relaxed) ? 1 : 0;
+}
+
+#else
+// Non-Android fallback — should not be reached; oboe_stub.cpp is used instead.
+// Provide empty implementations just in case.
+
+int wzp_oboe_start(const WzpOboeConfig* config, const WzpOboeRings* rings) {
+    (void)config; (void)rings;
+    return -99;
+}
+
+void wzp_oboe_stop(void) {}
+float wzp_oboe_capture_latency_ms(void) { return 0.0f; }
+float wzp_oboe_playout_latency_ms(void) { return 0.0f; }
+int wzp_oboe_is_running(void) { return 0; }
+
+#endif // __ANDROID__

crates/wzp-native/cpp/oboe_bridge.h

@@ -0,0 +1,44 @@
+#ifndef WZP_OBOE_BRIDGE_H
+#define WZP_OBOE_BRIDGE_H
+
+#include <stdint.h>
+
+#ifdef __cplusplus
+#include <atomic>
+typedef std::atomic<int32_t> wzp_atomic_int;
+extern "C" {
+#else
+#include <stdatomic.h>
+typedef atomic_int wzp_atomic_int;
+#endif
+
+typedef struct {
+    int32_t sample_rate;
+    int32_t frames_per_burst;
+    int32_t channel_count;
+    int32_t bt_active;  /* nonzero = BT SCO mode: skip sample rate + input preset */
+} WzpOboeConfig;
+
+typedef struct {
+    int16_t* capture_buf;
+    int32_t  capture_capacity;
+    wzp_atomic_int* capture_write_idx;
+    wzp_atomic_int* capture_read_idx;
+
+    int16_t* playout_buf;
+    int32_t  playout_capacity;
+    wzp_atomic_int* playout_write_idx;
+    wzp_atomic_int* playout_read_idx;
+} WzpOboeRings;
+
+int wzp_oboe_start(const WzpOboeConfig* config, const WzpOboeRings* rings);
+void wzp_oboe_stop(void);
+float wzp_oboe_capture_latency_ms(void);
+float wzp_oboe_playout_latency_ms(void);
+int wzp_oboe_is_running(void);
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif // WZP_OBOE_BRIDGE_H

crates/wzp-native/cpp/oboe_stub.cpp

@@ -0,0 +1,27 @@
+// Stub implementation for non-Android host builds (testing, cargo check, etc.)
+
+#include "oboe_bridge.h"
+#include <stdio.h>
+
+int wzp_oboe_start(const WzpOboeConfig* config, const WzpOboeRings* rings) {
+    (void)config;
+    (void)rings;
+    fprintf(stderr, "wzp_oboe_start: stub (not on Android)\n");
+    return 0;
+}
+
+void wzp_oboe_stop(void) {
+    fprintf(stderr, "wzp_oboe_stop: stub (not on Android)\n");
+}
+
+float wzp_oboe_capture_latency_ms(void) {
+    return 0.0f;
+}
+
+float wzp_oboe_playout_latency_ms(void) {
+    return 0.0f;
+}
+
+int wzp_oboe_is_running(void) {
+    return 0;
+}

crates/wzp-native/src/lib.rs

@@ -0,0 +1,449 @@
+//! wzp-native — standalone Android cdylib for all the C++ audio code.
+//!
+//! Built with `cargo ndk`, NOT `cargo tauri android build`. Loaded at
+//! runtime by the Tauri desktop cdylib (`wzp-desktop`) via libloading.
+//! See `docs/incident-tauri-android-init-tcb.md` for why the split exists.
+//!
+//! Phase 2: real Oboe audio backend.
+//!
+//! Architecture: Oboe runs capture + playout streams on its own high-
+//! priority AAudio callback threads inside the C++ bridge. Two SPSC ring
+//! buffers (capture and playout) are shared between the C++ callbacks
+//! and the Rust side via atomic indices — no locks on the hot path.
+//! `wzp-desktop` drains the capture ring into its Opus encoder and fills
+//! the playout ring with decoded PCM.
+
+use std::sync::atomic::{AtomicI32, Ordering};
+
+// ─── Phase 1 smoke-test exports (kept for sanity checks) ─────────────────
+
+/// Returns 42. Used by wzp-desktop's setup() to verify dlopen + dlsym
+/// work before any audio code runs.
+#[unsafe(no_mangle)]
+pub extern "C" fn wzp_native_version() -> i32 {
+    42
+}
+
+/// Writes a NUL-terminated string into `out` (capped at `cap`) and
+/// returns bytes written excluding the NUL.
+///
+/// # Safety
+/// `out` must be a valid pointer to at least `cap` contiguous bytes of
+/// writable memory. Passing a null pointer or zero capacity is safe
+/// (returns 0), but a dangling non-null pointer is undefined behaviour.
+#[unsafe(no_mangle)]
+pub unsafe extern "C" fn wzp_native_hello(out: *mut u8, cap: usize) -> usize {
+    const MSG: &[u8] = b"hello from wzp-native\0";
+    if out.is_null() || cap == 0 {
+        return 0;
+    }
+    let n = MSG.len().min(cap);
+    unsafe {
+        core::ptr::copy_nonoverlapping(MSG.as_ptr(), out, n);
+        *out.add(n - 1) = 0;
+    }
+    n - 1
+}
+
+// ─── C++ Oboe bridge FFI ─────────────────────────────────────────────────
+
+#[repr(C)]
+struct WzpOboeConfig {
+    sample_rate: i32,
+    frames_per_burst: i32,
+    channel_count: i32,
+    /// When nonzero, capture stream skips setSampleRate and setInputPreset
+    /// so the system can route to BT SCO at its native rate (8/16kHz).
+    /// Oboe's SampleRateConversionQuality::Best resamples to 48kHz.
+    bt_active: i32,
+}
+
+#[repr(C)]
+struct WzpOboeRings {
+    capture_buf: *mut i16,
+    capture_capacity: i32,
+    capture_write_idx: *mut AtomicI32,
+    capture_read_idx: *mut AtomicI32,
+    playout_buf: *mut i16,
+    playout_capacity: i32,
+    playout_write_idx: *mut AtomicI32,
+    playout_read_idx: *mut AtomicI32,
+}
+
+// SAFETY: atomics synchronise producer/consumer; raw pointers are owned
+// by the AudioBackend singleton below whose lifetime covers all calls.
+unsafe impl Send for WzpOboeRings {}
+unsafe impl Sync for WzpOboeRings {}
+
+unsafe extern "C" {
+    fn wzp_oboe_start(config: *const WzpOboeConfig, rings: *const WzpOboeRings) -> i32;
+    fn wzp_oboe_stop();
+    fn wzp_oboe_capture_latency_ms() -> f32;
+    fn wzp_oboe_playout_latency_ms() -> f32;
+    fn wzp_oboe_is_running() -> i32;
+}
+
+// ─── SPSC ring buffer (shared with C++ via AtomicI32) ────────────────────
+
+/// 20 ms @ 48 kHz mono = 960 samples.
+const FRAME_SAMPLES: usize = 960;
+/// ~160 ms headroom at 48 kHz.
+const RING_CAPACITY: usize = 7680;
+
+struct RingBuffer {
+    buf: Vec<i16>,
+    capacity: usize,
+    write_idx: AtomicI32,
+    read_idx: AtomicI32,
+}
+
+// SAFETY: SPSC with atomic read/write cursors; producer and consumer
+// are always on different threads.
+unsafe impl Send for RingBuffer {}
+unsafe impl Sync for RingBuffer {}
+
+impl RingBuffer {
+    fn new(capacity: usize) -> Self {
+        Self {
+            buf: vec![0i16; capacity],
+            capacity,
+            write_idx: AtomicI32::new(0),
+            read_idx: AtomicI32::new(0),
+        }
+    }
+
+    fn available_read(&self) -> usize {
+        let w = self.write_idx.load(Ordering::Acquire);
+        let r = self.read_idx.load(Ordering::Relaxed);
+        let avail = w - r;
+        if avail < 0 { (avail + self.capacity as i32) as usize } else { avail as usize }
+    }
+
+    fn available_write(&self) -> usize {
+        self.capacity - 1 - self.available_read()
+    }
+
+    fn write(&self, data: &[i16]) -> usize {
+        let count = data.len().min(self.available_write());
+        if count == 0 {
+            return 0;
+        }
+        let mut w = self.write_idx.load(Ordering::Relaxed) as usize;
+        let cap = self.capacity;
+        let buf_ptr = self.buf.as_ptr() as *mut i16;
+        for sample in &data[..count] {
+            unsafe { *buf_ptr.add(w) = *sample; }
+            w += 1;
+            if w >= cap { w = 0; }
+        }
+        self.write_idx.store(w as i32, Ordering::Release);
+        count
+    }
+
+    fn read(&self, out: &mut [i16]) -> usize {
+        let count = out.len().min(self.available_read());
+        if count == 0 {
+            return 0;
+        }
+        let mut r = self.read_idx.load(Ordering::Relaxed) as usize;
+        let cap = self.capacity;
+        let buf_ptr = self.buf.as_ptr();
+        for slot in &mut out[..count] {
+            unsafe { *slot = *buf_ptr.add(r); }
+            r += 1;
+            if r >= cap { r = 0; }
+        }
+        self.read_idx.store(r as i32, Ordering::Release);
+        count
+    }
+
+    fn buf_ptr(&self) -> *mut i16 {
+        self.buf.as_ptr() as *mut i16
+    }
+    fn write_idx_ptr(&self) -> *mut AtomicI32 {
+        &self.write_idx as *const AtomicI32 as *mut AtomicI32
+    }
+    fn read_idx_ptr(&self) -> *mut AtomicI32 {
+        &self.read_idx as *const AtomicI32 as *mut AtomicI32
+    }
+}
+
+// ─── AudioBackend singleton ──────────────────────────────────────────────
+//
+// There is one global AudioBackend instance because Oboe's C++ side
+// holds its own singleton of the streams. The `Box::leak`'d statics own
+// the ring buffers for the lifetime of the process — dropping them while
+// Oboe is still running would cause use-after-free in the audio callback.
+
+use std::sync::OnceLock;
+
+struct AudioBackend {
+    capture: RingBuffer,
+    playout: RingBuffer,
+    started: std::sync::Mutex<bool>,
+    /// Per-write logging throttle counter for wzp_native_audio_write_playout.
+    playout_write_log_count: std::sync::atomic::AtomicU64,
+    /// Fix A (task #35): the playout ring's read_idx at the last
+    /// check. If audio_write_playout observes read_idx hasn't
+    /// advanced after N writes, the Oboe playout callback has
+    /// stopped firing → restart the streams.
+    playout_last_read_idx: std::sync::atomic::AtomicI32,
+    /// Number of writes since the last read_idx advance.
+    playout_stall_writes: std::sync::atomic::AtomicU32,
+}
+
+static BACKEND: OnceLock<&'static AudioBackend> = OnceLock::new();
+
+fn backend() -> &'static AudioBackend {
+    BACKEND.get_or_init(|| {
+        Box::leak(Box::new(AudioBackend {
+            capture: RingBuffer::new(RING_CAPACITY),
+            playout: RingBuffer::new(RING_CAPACITY),
+            started: std::sync::Mutex::new(false),
+            playout_write_log_count: std::sync::atomic::AtomicU64::new(0),
+            playout_last_read_idx: std::sync::atomic::AtomicI32::new(0),
+            playout_stall_writes: std::sync::atomic::AtomicU32::new(0),
+        }))
+    })
+}
+
+// ─── C FFI for wzp-desktop ───────────────────────────────────────────────
+
+/// Start the Oboe audio streams. Returns 0 on success, non-zero on error.
+/// Idempotent — calling while already running is a no-op that returns 0.
+#[unsafe(no_mangle)]
+pub extern "C" fn wzp_native_audio_start() -> i32 {
+    audio_start_inner(false)
+}
+
+/// Start Oboe in Bluetooth SCO mode — skips sample rate and input preset
+/// on capture so the system can route to the BT SCO device natively.
+#[unsafe(no_mangle)]
+pub extern "C" fn wzp_native_audio_start_bt() -> i32 {
+    audio_start_inner(true)
+}
+
+fn audio_start_inner(bt: bool) -> i32 {
+    let b = backend();
+    let mut started = match b.started.lock() {
+        Ok(g) => g,
+        Err(_) => return -1,
+    };
+    if *started {
+        return 0;
+    }
+
+    let config = WzpOboeConfig {
+        sample_rate: 48_000,
+        frames_per_burst: FRAME_SAMPLES as i32,
+        channel_count: 1,
+        bt_active: if bt { 1 } else { 0 },
+    };
+    let rings = WzpOboeRings {
+        capture_buf: b.capture.buf_ptr(),
+        capture_capacity: b.capture.capacity as i32,
+        capture_write_idx: b.capture.write_idx_ptr(),
+        capture_read_idx: b.capture.read_idx_ptr(),
+        playout_buf: b.playout.buf_ptr(),
+        playout_capacity: b.playout.capacity as i32,
+        playout_write_idx: b.playout.write_idx_ptr(),
+        playout_read_idx: b.playout.read_idx_ptr(),
+    };
+    let ret = unsafe { wzp_oboe_start(&config, &rings) };
+    if ret != 0 {
+        return ret;
+    }
+    *started = true;
+    0
+}
+
+/// Stop Oboe. Idempotent. Safe to call from any thread.
+#[unsafe(no_mangle)]
+pub extern "C" fn wzp_native_audio_stop() {
+    let b = backend();
+    if let Ok(mut started) = b.started.lock() {
+        if *started {
+            unsafe { wzp_oboe_stop() };
+            *started = false;
+        }
+    }
+}
+
+/// Number of capture samples available to read without blocking.
+#[unsafe(no_mangle)]
+pub extern "C" fn wzp_native_audio_capture_available() -> usize {
+    backend().capture.available_read()
+}
+
+/// Read captured PCM samples from the capture ring. Returns the number
+/// of `i16` samples actually copied into `out` (may be less than
+/// `out_len` if the ring is empty).
+///
+/// # Safety
+/// `out` must be a valid pointer to `out_len` contiguous `i16` values.
+/// The caller must ensure no other thread writes to the same buffer
+/// concurrently. Passing a null pointer or zero length is safe (returns 0).
+#[unsafe(no_mangle)]
+pub unsafe extern "C" fn wzp_native_audio_read_capture(out: *mut i16, out_len: usize) -> usize {
+    if out.is_null() || out_len == 0 {
+        return 0;
+    }
+    let slice = unsafe { std::slice::from_raw_parts_mut(out, out_len) };
+    backend().capture.read(slice)
+}
+
+/// Write PCM samples into the playout ring. Returns the number of
+/// samples actually enqueued (may be less than `in_len` if the ring
+/// is nearly full — in practice the caller should pace to 20 ms
+/// frames and spin briefly if the ring is full).
+///
+/// # Safety
+/// `input` must be a valid pointer to `in_len` contiguous `i16` values
+/// that remain valid for the duration of the call. Passing a null pointer
+/// or zero length is safe (returns 0). The caller must not free or mutate
+/// the buffer while this function is executing.
+#[unsafe(no_mangle)]
+pub unsafe extern "C" fn wzp_native_audio_write_playout(input: *const i16, in_len: usize) -> usize {
+    if input.is_null() || in_len == 0 {
+        return 0;
+    }
+    let slice = unsafe { std::slice::from_raw_parts(input, in_len) };
+    let b = backend();
+
+    // Fix A (task #35): detect playout callback stall. If the
+    // playout ring's read_idx hasn't advanced in 50+ writes
+    // (~1 second at 50 writes/sec), the Oboe playout callback
+    // has stopped firing → restart the streams. This is the
+    // self-healing behavior that makes rejoin work: teardown +
+    // rebuild clears whatever HAL state locked up the callback.
+    let current_read_idx = b.playout.read_idx.load(std::sync::atomic::Ordering::Relaxed);
+    let last_read_idx = b.playout_last_read_idx.load(std::sync::atomic::Ordering::Relaxed);
+    if current_read_idx == last_read_idx {
+        let stall = b.playout_stall_writes.fetch_add(1, std::sync::atomic::Ordering::Relaxed);
+        if stall >= 50 {
+            // Callback hasn't drained anything in ~1 second.
+            // Force a stream restart.
+            unsafe {
+                android_log("playout STALL detected (50 writes, read_idx unchanged) — restarting Oboe streams");
+            }
+            b.playout_stall_writes.store(0, std::sync::atomic::Ordering::Relaxed);
+            // Release the started lock, stop, re-start.
+            // This is the same logic as the Rust-side
+            // audio_stop() + audio_start() but done inline
+            // because we can't call the extern "C" fns
+            // recursively. Just call the C++ side directly.
+            {
+                if let Ok(mut started) = b.started.lock() {
+                    if *started {
+                        unsafe { wzp_oboe_stop() };
+                        *started = false;
+                    }
+                }
+            }
+            // Clear the rings so the restart doesn't read stale data
+            b.playout.write_idx.store(0, std::sync::atomic::Ordering::Relaxed);
+            b.playout.read_idx.store(0, std::sync::atomic::Ordering::Relaxed);
+            b.capture.write_idx.store(0, std::sync::atomic::Ordering::Relaxed);
+            b.capture.read_idx.store(0, std::sync::atomic::Ordering::Relaxed);
+            // Re-start (stall detector — always non-BT mode)
+            let config = WzpOboeConfig {
+                sample_rate: 48_000,
+                frames_per_burst: FRAME_SAMPLES as i32,
+                channel_count: 1,
+                bt_active: 0,
+            };
+            let rings = WzpOboeRings {
+                capture_buf: b.capture.buf_ptr(),
+                capture_capacity: b.capture.capacity as i32,
+                capture_write_idx: b.capture.write_idx_ptr(),
+                capture_read_idx: b.capture.read_idx_ptr(),
+                playout_buf: b.playout.buf_ptr(),
+                playout_capacity: b.playout.capacity as i32,
+                playout_write_idx: b.playout.write_idx_ptr(),
+                playout_read_idx: b.playout.read_idx_ptr(),
+            };
+            let ret = unsafe { wzp_oboe_start(&config, &rings) };
+            if ret == 0 {
+                if let Ok(mut started) = b.started.lock() {
+                    *started = true;
+                }
+                unsafe { android_log("playout restart OK — Oboe streams rebuilt"); }
+            } else {
+                unsafe { android_log(&format!("playout restart FAILED: {ret}")); }
+            }
+            b.playout_last_read_idx.store(0, std::sync::atomic::Ordering::Relaxed);
+            return 0; // caller will retry on next frame
+        }
+    } else {
+        // read_idx advanced — callback is alive, reset counter
+        b.playout_stall_writes.store(0, std::sync::atomic::Ordering::Relaxed);
+        b.playout_last_read_idx.store(current_read_idx, std::sync::atomic::Ordering::Relaxed);
+    }
+
+    let before_w = b.playout.write_idx.load(std::sync::atomic::Ordering::Relaxed);
+    let before_r = b.playout.read_idx.load(std::sync::atomic::Ordering::Relaxed);
+    let written = b.playout.write(slice);
+    // First few writes: log ring state + sample range so we can compare what
+    // engine.rs hands us to what the C++ playout callback reads.
+    let first_writes = b.playout_write_log_count.fetch_add(1, std::sync::atomic::Ordering::Relaxed);
+    if first_writes < 3 || first_writes % 50 == 0 {
+        let (mut lo, mut hi, mut sumsq) = (i16::MAX, i16::MIN, 0i64);
+        for &s in slice.iter() {
+            if s < lo { lo = s; }
+            if s > hi { hi = s; }
+            sumsq += (s as i64) * (s as i64);
+        }
+        let rms = (sumsq as f64 / slice.len() as f64).sqrt() as i32;
+        let avail_w_after = b.playout.available_write();
+        let avail_r_after = b.playout.available_read();
+        let msg = format!(
+            "playout WRITE #{first_writes}: in_len={} written={} range=[{lo}..{hi}] rms={rms} before_w={before_w} before_r={before_r} avail_read_after={avail_r_after} avail_write_after={avail_w_after}",
+            slice.len(), written
+        );
+        unsafe {
+            android_log(msg.as_str());
+        }
+    }
+    written
+}
+
+// Minimal android logcat shim so we can print from the cdylib without pulling
+// in android_logger crate (which would add another dep that has to build with
+// cargo-ndk). Uses libc's __android_log_print via extern linkage.
+#[cfg(target_os = "android")]
+unsafe extern "C" {
+    fn __android_log_write(prio: i32, tag: *const u8, text: *const u8) -> i32;
+}
+
+#[cfg(target_os = "android")]
+unsafe fn android_log(msg: &str) {
+    // ANDROID_LOG_INFO = 4. Tag and text must be NUL-terminated.
+    let tag = b"wzp-native\0";
+    let mut buf = Vec::with_capacity(msg.len() + 1);
+    buf.extend_from_slice(msg.as_bytes());
+    buf.push(0);
+    unsafe { __android_log_write(4, tag.as_ptr(), buf.as_ptr()); }
+}
+
+#[cfg(not(target_os = "android"))]
+#[allow(dead_code)]
+unsafe fn android_log(_msg: &str) {}
+
+/// Current capture latency reported by Oboe, in milliseconds. Returns
+/// NaN / 0.0 if the stream isn't running.
+#[unsafe(no_mangle)]
+pub extern "C" fn wzp_native_audio_capture_latency_ms() -> f32 {
+    unsafe { wzp_oboe_capture_latency_ms() }
+}
+
+/// Current playout latency reported by Oboe, in milliseconds.
+#[unsafe(no_mangle)]
+pub extern "C" fn wzp_native_audio_playout_latency_ms() -> f32 {
+    unsafe { wzp_oboe_playout_latency_ms() }
+}
+
+/// Non-zero if both Oboe streams are currently running.
+#[unsafe(no_mangle)]
+pub extern "C" fn wzp_native_audio_is_running() -> i32 {
+    unsafe { wzp_oboe_is_running() }
+}

crates/wzp-proto/src/dred_tuner.rs

@@ -0,0 +1,316 @@
+//! Continuous DRED tuning from real-time network metrics.
+//!
+//! Instead of locking DRED duration to 3 discrete quality tiers (100/200/500 ms),
+//! `DredTuner` maps live path quality metrics to a continuous DRED duration and
+//! expected-loss hint, updated every N packets. This makes DRED reactive within
+//! ~200 ms instead of waiting for 3+ consecutive bad quality reports to trigger
+//! a full tier transition.
+//!
+//! The tuner also implements pre-emptive jitter-spike detection ("sawtooth"
+//! prediction): when jitter variance spikes >30% over a 200 ms window — typical
+//! of Starlink satellite handovers — it temporarily boosts DRED to the maximum
+//! allowed for the current codec before packets actually start dropping.
+//!
+//! See also: [`crate::quality`] for discrete tier classification that drives
+//! codec switching. DredTuner operates within a tier, adjusting DRED
+//! parameters continuously based on live network metrics.
+
+use crate::CodecId;
+
+/// Output of a single tuning cycle.
+#[derive(Clone, Copy, Debug, PartialEq, Eq)]
+pub struct DredTuning {
+    /// DRED duration in 10 ms frame units (0–104). Passed directly to
+    /// `OpusEncoder::set_dred_duration()`.
+    pub dred_frames: u8,
+    /// Expected packet loss percentage (0–100). Passed to
+    /// `OpusEncoder::set_expected_loss()`. Floored at 15% by the encoder
+    /// itself, but we pass the real value so the encoder can override upward.
+    pub expected_loss_pct: u8,
+}
+
+/// Minimum DRED frames for any Opus codec (matches DRED_LOSS_FLOOR_PCT logic:
+/// at 15% loss, libopus 1.5 emits ~95 ms of DRED, which needs at least 10
+/// frames configured to be useful).
+const MIN_DRED_FRAMES: u8 = 5;
+
+/// Maximum DRED frames libopus supports (104 × 10 ms = 1040 ms).
+const MAX_DRED_FRAMES: u8 = 104;
+
+/// Jitter variance spike ratio that triggers pre-emptive DRED boost.
+const JITTER_SPIKE_RATIO: f32 = 1.3;
+
+/// How many tuning cycles a jitter-spike boost persists (at 25 packets/cycle
+/// and 20 ms/packet, 10 cycles ≈ 5 seconds).
+const SPIKE_BOOST_COOLDOWN_CYCLES: u32 = 10;
+
+/// Maps codec tier to its baseline DRED frames (used when network is healthy).
+fn baseline_dred_frames(codec: CodecId) -> u8 {
+    match codec {
+        CodecId::Opus32k | CodecId::Opus48k | CodecId::Opus64k => 10, // 100 ms
+        CodecId::Opus16k | CodecId::Opus24k => 20,                    // 200 ms
+        CodecId::Opus6k => 50,                                         // 500 ms
+        _ => 0,
+    }
+}
+
+/// Maps codec tier to its maximum allowed DRED frames under spike/bad conditions.
+fn max_dred_frames_for(codec: CodecId) -> u8 {
+    match codec {
+        // Studio: cap at 300 ms (don't waste bitrate on good links)
+        CodecId::Opus32k | CodecId::Opus48k | CodecId::Opus64k => 30,
+        // Normal: cap at 500 ms
+        CodecId::Opus16k | CodecId::Opus24k => 50,
+        // Degraded: allow full 1040 ms
+        CodecId::Opus6k => MAX_DRED_FRAMES,
+        _ => 0,
+    }
+}
+
+/// Continuous DRED tuner driven by network path metrics.
+pub struct DredTuner {
+    /// Current codec (determines baseline and ceiling).
+    codec: CodecId,
+    /// Last computed tuning output.
+    last_tuning: DredTuning,
+    /// EWMA-smoothed jitter for spike detection (in ms).
+    jitter_ewma: f32,
+    /// Remaining cooldown cycles for a jitter-spike boost.
+    spike_cooldown: u32,
+    /// Whether the tuner has received at least one observation.
+    initialized: bool,
+}
+
+impl DredTuner {
+    /// Create a new tuner for the given codec.
+    pub fn new(codec: CodecId) -> Self {
+        let baseline = baseline_dred_frames(codec);
+        Self {
+            codec,
+            last_tuning: DredTuning {
+                dred_frames: baseline,
+                expected_loss_pct: 15, // match DRED_LOSS_FLOOR_PCT
+            },
+            jitter_ewma: 0.0,
+            spike_cooldown: 0,
+            initialized: false,
+        }
+    }
+
+    /// Update the active codec (e.g. on tier transition). Resets spike state.
+    pub fn set_codec(&mut self, codec: CodecId) {
+        self.codec = codec;
+        self.spike_cooldown = 0;
+    }
+
+    /// Feed network metrics and compute new DRED parameters.
+    ///
+    /// Call this every tuning cycle (e.g. every 25 packets ≈ 500 ms at 20 ms
+    /// frame duration).
+    ///
+    /// - `loss_pct`: observed packet loss (0.0–100.0)
+    /// - `rtt_ms`: smoothed round-trip time
+    /// - `jitter_ms`: current jitter estimate (RTT variance)
+    ///
+    /// Returns `Some(tuning)` if the output changed, `None` if unchanged.
+    pub fn update(&mut self, loss_pct: f32, rtt_ms: u32, jitter_ms: u32) -> Option<DredTuning> {
+        if !self.codec.is_opus() {
+            return None;
+        }
+
+        let baseline = baseline_dred_frames(self.codec);
+        let ceiling = max_dred_frames_for(self.codec);
+
+        // --- Jitter spike detection ---
+        let jitter_f = jitter_ms as f32;
+        if !self.initialized {
+            self.jitter_ewma = jitter_f;
+            self.initialized = true;
+        } else {
+            // Fast-up (alpha=0.3), slow-down (alpha=0.05) asymmetric EWMA
+            let alpha = if jitter_f > self.jitter_ewma { 0.3 } else { 0.05 };
+            self.jitter_ewma = alpha * jitter_f + (1.0 - alpha) * self.jitter_ewma;
+        }
+
+        // Detect spike: instantaneous jitter > EWMA × 1.3
+        if self.jitter_ewma > 1.0 && jitter_f > self.jitter_ewma * JITTER_SPIKE_RATIO {
+            self.spike_cooldown = SPIKE_BOOST_COOLDOWN_CYCLES;
+        }
+
+        // Decrement cooldown
+        if self.spike_cooldown > 0 {
+            self.spike_cooldown -= 1;
+        }
+
+        // --- Compute DRED frames ---
+        let dred_frames = if self.spike_cooldown > 0 {
+            // During spike boost: jump to ceiling
+            ceiling
+        } else {
+            // Continuous mapping: scale linearly between baseline and ceiling
+            // based on loss percentage.
+            //   0% loss → baseline
+            //  40% loss → ceiling
+            let loss_clamped = loss_pct.clamp(0.0, 40.0);
+            let t = loss_clamped / 40.0;
+            let raw = baseline as f32 + t * (ceiling - baseline) as f32;
+            (raw as u8).clamp(MIN_DRED_FRAMES, ceiling)
+        };
+
+        // --- Compute expected loss hint ---
+        // Pass the real loss so the encoder can clamp at its own floor (15%).
+        // For RTT-driven boost: high RTT suggests impending loss, so add a
+        // phantom loss contribution to keep DRED emitting generously.
+        let rtt_loss_phantom = if rtt_ms > 200 {
+            ((rtt_ms - 200) as f32 / 40.0).min(15.0)
+        } else {
+            0.0
+        };
+        let expected_loss = (loss_pct + rtt_loss_phantom).clamp(0.0, 100.0) as u8;
+
+        let tuning = DredTuning {
+            dred_frames,
+            expected_loss_pct: expected_loss,
+        };
+
+        if tuning != self.last_tuning {
+            self.last_tuning = tuning;
+            Some(tuning)
+        } else {
+            None
+        }
+    }
+
+    /// Get the last computed tuning without updating.
+    pub fn current(&self) -> DredTuning {
+        self.last_tuning
+    }
+
+    /// Whether a jitter-spike boost is currently active.
+    pub fn spike_boost_active(&self) -> bool {
+        self.spike_cooldown > 0
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn baseline_for_opus24k() {
+        let tuner = DredTuner::new(CodecId::Opus24k);
+        assert_eq!(tuner.current().dred_frames, 20); // 200 ms
+    }
+
+    #[test]
+    fn baseline_for_opus6k() {
+        let tuner = DredTuner::new(CodecId::Opus6k);
+        assert_eq!(tuner.current().dred_frames, 50); // 500 ms
+    }
+
+    #[test]
+    fn codec2_returns_none() {
+        let mut tuner = DredTuner::new(CodecId::Codec2_1200);
+        assert!(tuner.update(10.0, 100, 20).is_none());
+    }
+
+    #[test]
+    fn scales_with_loss() {
+        let mut tuner = DredTuner::new(CodecId::Opus24k);
+
+        // 0% loss → baseline (20 frames)
+        tuner.update(0.0, 50, 5);
+        assert_eq!(tuner.current().dred_frames, 20);
+
+        // 20% loss → midpoint between 20 and 50 = 35
+        tuner.update(20.0, 50, 5);
+        assert_eq!(tuner.current().dred_frames, 35);
+
+        // 40%+ loss → ceiling (50 frames)
+        tuner.update(40.0, 50, 5);
+        assert_eq!(tuner.current().dred_frames, 50);
+    }
+
+    #[test]
+    fn jitter_spike_triggers_boost() {
+        let mut tuner = DredTuner::new(CodecId::Opus24k);
+
+        // Establish baseline jitter
+        for _ in 0..20 {
+            tuner.update(0.0, 50, 10);
+        }
+        assert!(!tuner.spike_boost_active());
+
+        // Spike: jitter jumps to 50 ms (5x the EWMA of ~10)
+        tuner.update(0.0, 50, 50);
+        assert!(tuner.spike_boost_active());
+        // Should be at ceiling (50 frames = 500 ms for Opus24k)
+        assert_eq!(tuner.current().dred_frames, 50);
+    }
+
+    #[test]
+    fn spike_cooldown_decays() {
+        let mut tuner = DredTuner::new(CodecId::Opus24k);
+
+        // Establish baseline then spike
+        for _ in 0..20 {
+            tuner.update(0.0, 50, 10);
+        }
+        tuner.update(0.0, 50, 50);
+        assert!(tuner.spike_boost_active());
+
+        // Run through cooldown
+        for _ in 0..SPIKE_BOOST_COOLDOWN_CYCLES {
+            tuner.update(0.0, 50, 10);
+        }
+        assert!(!tuner.spike_boost_active());
+        // Should return to baseline
+        assert_eq!(tuner.current().dred_frames, 20);
+    }
+
+    #[test]
+    fn rtt_phantom_loss() {
+        let mut tuner = DredTuner::new(CodecId::Opus24k);
+
+        // High RTT (400ms) with 0% real loss
+        tuner.update(0.0, 400, 10);
+        // Phantom loss = (400-200)/40 = 5
+        assert_eq!(tuner.current().expected_loss_pct, 5);
+    }
+
+    #[test]
+    fn set_codec_resets_spike() {
+        let mut tuner = DredTuner::new(CodecId::Opus24k);
+
+        // Trigger spike
+        for _ in 0..20 {
+            tuner.update(0.0, 50, 10);
+        }
+        tuner.update(0.0, 50, 50);
+        assert!(tuner.spike_boost_active());
+
+        // Switch codec — spike should reset
+        tuner.set_codec(CodecId::Opus6k);
+        assert!(!tuner.spike_boost_active());
+    }
+
+    #[test]
+    fn opus6k_reaches_max_1040ms() {
+        let mut tuner = DredTuner::new(CodecId::Opus6k);
+
+        // High loss → should reach 104 frames (1040 ms)
+        tuner.update(40.0, 50, 5);
+        assert_eq!(tuner.current().dred_frames, MAX_DRED_FRAMES);
+    }
+
+    #[test]
+    fn returns_none_when_unchanged() {
+        let mut tuner = DredTuner::new(CodecId::Opus24k);
+
+        // First update always returns Some (initial → computed)
+        let first = tuner.update(0.0, 50, 5);
+        // Same inputs → None
+        let second = tuner.update(0.0, 50, 5);
+        assert!(first.is_some() || second.is_none());
+    }
+}

crates/wzp-proto/src/error.rs

@@ -53,6 +53,15 @@ pub enum TransportError {
     Timeout { ms: u64 },
     #[error("io error: {0}")]
     Io(#[from] std::io::Error),
+    /// Parsed wire bytes successfully but the payload didn't
+    /// deserialize into a known `SignalMessage` variant. Usually
+    /// means the peer is running a newer build with a variant we
+    /// don't know yet. Callers should **log and continue** rather
+    /// than tearing down the connection, so that forward-compat
+    /// additions to `SignalMessage` don't silently kill old
+    /// clients/relays.
+    #[error("signal deserialize: {0}")]
+    Deserialize(String),
     #[error("internal transport error: {0}")]
     Internal(String),
 }

crates/wzp-proto/src/lib.rs

@@ -14,6 +14,7 @@
 
 pub mod bandwidth;
 pub mod codec_id;
+pub mod dred_tuner;
 pub mod error;
 pub mod jitter;
 pub mod packet;
@@ -26,10 +27,11 @@ pub use codec_id::{CodecId, QualityProfile};
 pub use error::*;
 pub use packet::{
     CallAcceptMode, HangupReason, MediaHeader, MediaPacket, MiniFrameContext, MiniHeader,
-    QualityReport, RoomParticipant, SignalMessage, TrunkEntry, TrunkFrame, FRAME_TYPE_FULL,
+    PresenceUser, QualityReport, RoomParticipant, SignalMessage, TrunkEntry, TrunkFrame, FRAME_TYPE_FULL,
     FRAME_TYPE_MINI,
 };
 pub use bandwidth::{BandwidthEstimator, CongestionState};
+pub use dred_tuner::{DredTuner, DredTuning};
 pub use quality::{AdaptiveQualityController, NetworkContext, Tier};
 pub use session::{Session, SessionEvent, SessionState};
 pub use traits::*;

crates/wzp-proto/src/quality.rs

@@ -1,3 +1,5 @@
+//! See also: [`crate::dred_tuner`] for continuous DRED tuning within a tier.
+
 use std::collections::VecDeque;
 use std::time::{Duration, Instant};
 
@@ -6,19 +8,31 @@ use crate::traits::QualityController;
 use crate::QualityProfile;
 
 /// Network quality tier — drives codec and FEC selection.
-#[derive(Clone, Copy, Debug, PartialEq, Eq)]
+///
+/// 5-tier range from studio quality down to catastrophic:
+/// Studio64k > Studio48k > Studio32k > Good > Degraded > Catastrophic
+#[derive(Clone, Copy, Debug, PartialEq, Eq, PartialOrd, Ord)]
 pub enum Tier {
-    /// loss < 10%, RTT < 400ms
-    Good,
-    /// loss 10-40% OR RTT 400-600ms
-    Degraded,
-    /// loss > 40% OR RTT > 600ms
-    Catastrophic,
+    /// loss >= 15% OR RTT >= 200ms — Codec2 1.2k
+    Catastrophic = 0,
+    /// loss < 15% AND RTT < 200ms — Opus 6k
+    Degraded = 1,
+    /// loss < 5% AND RTT < 100ms — Opus 24k
+    Good = 2,
+    /// loss < 2% AND RTT < 80ms — Opus 32k
+    Studio32k = 3,
+    /// loss < 1% AND RTT < 50ms — Opus 48k
+    Studio48k = 4,
+    /// loss < 1% AND RTT < 30ms — Opus 64k
+    Studio64k = 5,
 }
 
 impl Tier {
     pub fn profile(self) -> QualityProfile {
         match self {
+            Self::Studio64k => QualityProfile::STUDIO_64K,
+            Self::Studio48k => QualityProfile::STUDIO_48K,
+            Self::Studio32k => QualityProfile::STUDIO_32K,
             Self::Good => QualityProfile::GOOD,
             Self::Degraded => QualityProfile::DEGRADED,
             Self::Catastrophic => QualityProfile::CATASTROPHIC,
@@ -39,7 +53,7 @@ impl Tier {
             NetworkContext::CellularLte
             | NetworkContext::Cellular5g
             | NetworkContext::Cellular3g => {
-                // Tighter thresholds for cellular networks
+                // Tighter thresholds for cellular — no studio tiers
                 if loss > 25.0 || rtt > 500 {
                     Self::Catastrophic
                 } else if loss > 8.0 || rtt > 300 {
@@ -49,13 +63,18 @@ impl Tier {
                 }
             }
             NetworkContext::WiFi | NetworkContext::Unknown => {
-                // Original thresholds
-                if loss > 40.0 || rtt > 600 {
+                if loss >= 15.0 || rtt >= 200 {
                     Self::Catastrophic
-                } else if loss > 10.0 || rtt > 400 {
+                } else if loss >= 5.0 || rtt >= 100 {
                     Self::Degraded
-                } else {
+                } else if loss >= 2.0 || rtt >= 80 {
                     Self::Good
+                } else if loss >= 1.0 || rtt >= 50 {
+                    Self::Studio32k
+                } else if rtt >= 30 {
+                    Self::Studio48k
+                } else {
+                    Self::Studio64k
                 }
             }
         }
@@ -64,11 +83,19 @@ impl Tier {
     /// Return the next lower (worse) tier, or None if already at the worst.
     pub fn downgrade(self) -> Option<Tier> {
         match self {
+            Self::Studio64k => Some(Self::Studio48k),
+            Self::Studio48k => Some(Self::Studio32k),
+            Self::Studio32k => Some(Self::Good),
             Self::Good => Some(Self::Degraded),
             Self::Degraded => Some(Self::Catastrophic),
             Self::Catastrophic => None,
         }
     }
+
+    /// Whether this is a studio tier (above Good).
+    pub fn is_studio(self) -> bool {
+        matches!(self, Self::Studio64k | Self::Studio48k | Self::Studio32k)
+    }
 }
 
 /// Describes the network transport type for context-aware quality decisions.
@@ -108,20 +135,48 @@ pub struct AdaptiveQualityController {
     fec_boost_until: Option<Instant>,
     /// FEC boost amount to add during handoff recovery window.
     fec_boost_amount: f32,
+    /// Probing state: when Some, we're actively testing a higher tier.
+    probe: Option<ProbeState>,
+    /// Time spent stable at the current tier (for probe trigger).
+    stable_since: Option<Instant>,
 }
 
 /// Threshold for downgrading (fast reaction to degradation).
 const DOWNGRADE_THRESHOLD: u32 = 3;
 /// Threshold for downgrading on cellular networks (even faster).
 const CELLULAR_DOWNGRADE_THRESHOLD: u32 = 2;
-/// Threshold for upgrading (slow, cautious improvement).
-const UPGRADE_THRESHOLD: u32 = 10;
+/// Threshold for upgrading from Catastrophic/Degraded to Good.
+const UPGRADE_THRESHOLD: u32 = 5;
+/// Threshold for upgrading into studio tiers (very conservative).
+const STUDIO_UPGRADE_THRESHOLD: u32 = 10;
 /// Maximum history window size.
 const HISTORY_SIZE: usize = 20;
 /// Default FEC boost amount during handoff recovery.
 const DEFAULT_FEC_BOOST: f32 = 0.2;
 /// Duration of FEC boost after a network handoff.
 const FEC_BOOST_DURATION_SECS: u64 = 10;
+/// Minimum time stable at current tier before probing upward (30 seconds).
+const PROBE_STABLE_SECS: u64 = 30;
+/// Duration of a probe window (5 seconds — ~25 quality reports at 1/s).
+const PROBE_DURATION_SECS: u64 = 5;
+/// Maximum bad reports during probe before aborting (1 out of ~5 = 20%).
+const PROBE_MAX_BAD: u32 = 1;
+/// Cooldown after a failed probe before trying again (60 seconds).
+const PROBE_COOLDOWN_SECS: u64 = 60;
+
+/// Active bandwidth probe state.
+struct ProbeState {
+    /// The tier we're probing (one step above current).
+    target_tier: Tier,
+    /// Profile to apply during probe.
+    target_profile: QualityProfile,
+    /// When the probe started.
+    started: Instant,
+    /// Reports observed during probe.
+    probe_reports: u32,
+    /// Bad reports during probe (loss/RTT exceeded target tier thresholds).
+    bad_reports: u32,
+}
 
 impl AdaptiveQualityController {
     pub fn new() -> Self {
@@ -135,6 +190,8 @@ impl AdaptiveQualityController {
             network_context: NetworkContext::default(),
             fec_boost_until: None,
             fec_boost_amount: DEFAULT_FEC_BOOST,
+            probe: None,
+            stable_since: None,
         }
     }
 
@@ -174,6 +231,10 @@ impl AdaptiveQualityController {
             self.forced = false;
         }
 
+        // Cancel any active probe
+        self.probe = None;
+        self.stable_since = None;
+
         // Activate FEC boost for any network change
         self.fec_boost_until = Some(Instant::now() + Duration::from_secs(FEC_BOOST_DURATION_SECS));
     }
@@ -194,6 +255,8 @@ impl AdaptiveQualityController {
     pub fn reset_counters(&mut self) {
         self.consecutive_up = 0;
         self.consecutive_down = 0;
+        self.probe = None;
+        self.stable_since = None;
     }
 
     /// Get the effective downgrade threshold based on network context.
@@ -213,16 +276,13 @@ impl AdaptiveQualityController {
             return None;
         }
 
-        let is_worse = match (self.current_tier, observed_tier) {
-            (Tier::Good, Tier::Degraded | Tier::Catastrophic) => true,
-            (Tier::Degraded, Tier::Catastrophic) => true,
-            _ => false,
-        };
+        let is_worse = observed_tier < self.current_tier;
 
         if is_worse {
             self.consecutive_up = 0;
             self.consecutive_down += 1;
             if self.consecutive_down >= self.downgrade_threshold() {
+                // Jump directly to the observed tier (don't step one-at-a-time on downgrade)
                 self.current_tier = observed_tier;
                 self.current_profile = observed_tier.profile();
                 self.consecutive_down = 0;
@@ -232,22 +292,115 @@ impl AdaptiveQualityController {
             // Better conditions
             self.consecutive_down = 0;
             self.consecutive_up += 1;
-            if self.consecutive_up >= UPGRADE_THRESHOLD {
+            // Studio tiers require more consecutive good reports
+            let threshold = if self.current_tier >= Tier::Good {
+                STUDIO_UPGRADE_THRESHOLD
+            } else {
+                UPGRADE_THRESHOLD
+            };
+            if self.consecutive_up >= threshold {
                 // Only upgrade one step at a time
-                let next_tier = match self.current_tier {
-                    Tier::Catastrophic => Tier::Degraded,
-                    Tier::Degraded => Tier::Good,
-                    Tier::Good => return None,
-                };
-                self.current_tier = next_tier;
-                self.current_profile = next_tier.profile();
-                self.consecutive_up = 0;
-                return Some(self.current_profile);
+                if let Some(next_tier) = self.upgrade_one_step() {
+                    self.current_tier = next_tier;
+                    self.current_profile = next_tier.profile();
+                    self.consecutive_up = 0;
+                    return Some(self.current_profile);
+                }
             }
         }
 
         None
     }
+
+    /// Check whether to start, continue, or conclude a bandwidth probe.
+    ///
+    /// Called from `observe()` when no hysteresis transition fired.
+    fn check_probe(&mut self, observed_tier: Tier) -> Option<QualityProfile> {
+        // Don't probe if forced, or if already at highest tier, or on cellular
+        if self.forced || self.current_tier == Tier::Studio64k {
+            return None;
+        }
+        if matches!(
+            self.network_context,
+            NetworkContext::CellularLte | NetworkContext::Cellular5g | NetworkContext::Cellular3g
+        ) {
+            return None;
+        }
+
+        // If we have an active probe, evaluate it
+        if let Some(ref mut probe) = self.probe {
+            probe.probe_reports += 1;
+
+            // Check if the observed tier meets the probe target
+            if observed_tier < probe.target_tier {
+                probe.bad_reports += 1;
+            }
+
+            // Probe failed: too many bad reports
+            if probe.bad_reports > PROBE_MAX_BAD {
+                let _failed_probe = self.probe.take();
+                // Reset stable_since to trigger cooldown
+                self.stable_since =
+                    Some(Instant::now() + Duration::from_secs(PROBE_COOLDOWN_SECS));
+                return None; // stay at current tier
+            }
+
+            // Probe succeeded: enough good reports within the window
+            if probe.started.elapsed() >= Duration::from_secs(PROBE_DURATION_SECS) {
+                let target = probe.target_tier;
+                let profile = probe.target_profile;
+                self.probe.take();
+                self.current_tier = target;
+                self.current_profile = profile;
+                self.consecutive_up = 0;
+                self.stable_since = Some(Instant::now());
+                return Some(profile);
+            }
+
+            return None; // probe still running
+        }
+
+        // No active probe — check if we should start one
+        if observed_tier >= self.current_tier {
+            // Track stability
+            if self.stable_since.is_none() {
+                self.stable_since = Some(Instant::now());
+            }
+
+            if let Some(stable_since) = self.stable_since {
+                if stable_since.elapsed() >= Duration::from_secs(PROBE_STABLE_SECS) {
+                    // Stable long enough — start probing
+                    if let Some(next) = self.upgrade_one_step() {
+                        self.probe = Some(ProbeState {
+                            target_tier: next,
+                            target_profile: next.profile(),
+                            started: Instant::now(),
+                            probe_reports: 0,
+                            bad_reports: 0,
+                        });
+                        // Return the probe profile so the encoder switches
+                        return Some(next.profile());
+                    }
+                }
+            }
+        } else {
+            // Conditions degraded — reset stability timer
+            self.stable_since = None;
+        }
+
+        None
+    }
+
+    fn upgrade_one_step(&self) -> Option<Tier> {
+        match self.current_tier {
+            Tier::Catastrophic => Some(Tier::Degraded),
+            Tier::Degraded => Some(Tier::Good),
+            Tier::Good => Some(Tier::Studio32k),
+            Tier::Studio32k => Some(Tier::Studio48k),
+            Tier::Studio48k => Some(Tier::Studio64k),
+            Tier::Studio64k => None,
+        }
+    }
 }
 
 impl Default for AdaptiveQualityController {
@@ -269,7 +422,17 @@ impl QualityController for AdaptiveQualityController {
         }
 
         let observed = Tier::classify_with_context(report, self.network_context);
-        self.try_transition(observed)
+
+        // First check for downgrades/upgrades via hysteresis
+        if let Some(profile) = self.try_transition(observed) {
+            // Cancel any active probe on tier change
+            self.probe.take();
+            self.stable_since = None;
+            return Some(profile);
+        }
+
+        // Then check probing
+        self.check_probe(observed)
     }
 
     fn force_profile(&mut self, profile: QualityProfile) {
@@ -331,25 +494,33 @@ mod tests {
         }
         assert_eq!(ctrl.tier(), Tier::Catastrophic);
 
-        // 9 good reports — not enough
-        let good = make_report(2.0, 100);
-        for _ in 0..9 {
+        // 4 good reports — not enough (threshold is 5)
+        let good = make_report(0.5, 20); // studio-quality report
+        for _ in 0..4 {
             assert!(ctrl.observe(&good).is_none());
         }
         assert_eq!(ctrl.tier(), Tier::Catastrophic);
 
-        // 10th good report triggers upgrade (one step: Catastrophic → Degraded)
+        // 5th good report triggers upgrade (one step: Catastrophic → Degraded)
         let result = ctrl.observe(&good);
         assert!(result.is_some());
         assert_eq!(ctrl.tier(), Tier::Degraded);
 
-        // Need another 10 to go from Degraded → Good
-        for _ in 0..9 {
+        // Another 5 to go from Degraded → Good
+        for _ in 0..4 {
             assert!(ctrl.observe(&good).is_none());
         }
         let result = ctrl.observe(&good);
         assert!(result.is_some());
         assert_eq!(ctrl.tier(), Tier::Good);
+
+        // Studio upgrades need 10 consecutive — Good → Studio32k
+        for _ in 0..9 {
+            assert!(ctrl.observe(&good).is_none());
+        }
+        let result = ctrl.observe(&good);
+        assert!(result.is_some());
+        assert_eq!(ctrl.tier(), Tier::Studio32k);
     }
 
     #[test]
@@ -366,11 +537,29 @@ mod tests {
 
     #[test]
     fn tier_classification() {
-        assert_eq!(Tier::classify(&make_report(5.0, 200)), Tier::Good);
-        assert_eq!(Tier::classify(&make_report(15.0, 200)), Tier::Degraded);
-        assert_eq!(Tier::classify(&make_report(5.0, 500)), Tier::Degraded);
-        assert_eq!(Tier::classify(&make_report(50.0, 200)), Tier::Catastrophic);
-        assert_eq!(Tier::classify(&make_report(5.0, 700)), Tier::Catastrophic);
+        // Studio tiers
+        assert_eq!(Tier::classify(&make_report(0.5, 20)), Tier::Studio64k);
+        assert_eq!(Tier::classify(&make_report(0.5, 40)), Tier::Studio48k);
+        assert_eq!(Tier::classify(&make_report(1.5, 60)), Tier::Studio32k);
+        // Good/Degraded/Catastrophic
+        assert_eq!(Tier::classify(&make_report(3.0, 90)), Tier::Good);
+        assert_eq!(Tier::classify(&make_report(6.0, 120)), Tier::Degraded);
+        assert_eq!(Tier::classify(&make_report(16.0, 120)), Tier::Catastrophic);
+        assert_eq!(Tier::classify(&make_report(5.0, 200)), Tier::Catastrophic);
+    }
+
+    #[test]
+    fn studio_tier_boundaries() {
+        // loss < 1% AND RTT < 30ms → Studio64k
+        assert_eq!(Tier::classify(&make_report(0.9, 28)), Tier::Studio64k);
+        // loss < 1% AND RTT 30-49ms → Studio48k
+        assert_eq!(Tier::classify(&make_report(0.9, 32)), Tier::Studio48k);
+        // loss < 2% AND RTT < 80ms → Studio32k (but loss >= 1%)
+        assert_eq!(Tier::classify(&make_report(1.5, 40)), Tier::Studio32k);
+        // loss >= 2% → Good (use 2.5 to survive u8 quantization)
+        assert_eq!(Tier::classify(&make_report(2.5, 40)), Tier::Good);
+        // RTT 80ms → Good
+        assert_eq!(Tier::classify(&make_report(0.5, 80)), Tier::Good);
     }
 
     // ---------------------------------------------------------------
@@ -379,8 +568,8 @@ mod tests {
 
     #[test]
     fn cellular_tighter_thresholds() {
-        // 12% loss: Good on WiFi, Degraded on cellular
-        let report = make_report(12.0, 200);
+        // 9% loss: Degraded on both WiFi (>=5%) and cellular (>=8%)
+        let report = make_report(9.0, 80);
         assert_eq!(
             Tier::classify_with_context(&report, NetworkContext::WiFi),
             Tier::Degraded
@@ -390,22 +579,22 @@ mod tests {
             Tier::Degraded
         );
 
-        // 9% loss: Good on WiFi, Degraded on cellular
-        let report = make_report(9.0, 200);
+        // 6% loss, low RTT: Degraded on WiFi (>=5%), Good on cellular (<8%)
+        let report = make_report(6.0, 80);
         assert_eq!(
             Tier::classify_with_context(&report, NetworkContext::WiFi),
+            Tier::Degraded
+        );
+        assert_eq!(
+            Tier::classify_with_context(&report, NetworkContext::CellularLte),
             Tier::Good
         );
-        assert_eq!(
-            Tier::classify_with_context(&report, NetworkContext::CellularLte),
-            Tier::Degraded
-        );
 
-        // 30% loss: Degraded on WiFi, Catastrophic on cellular
-        let report = make_report(30.0, 200);
+        // 30% loss: Catastrophic on WiFi (>=15%), Catastrophic on cellular (>=25%)
+        let report = make_report(30.0, 80);
         assert_eq!(
             Tier::classify_with_context(&report, NetworkContext::WiFi),
-            Tier::Degraded
+            Tier::Catastrophic
         );
         assert_eq!(
             Tier::classify_with_context(&report, NetworkContext::Cellular3g),
@@ -415,15 +604,29 @@ mod tests {
 
     #[test]
     fn cellular_rtt_thresholds() {
-        // RTT 350ms: Good on WiFi, Degraded on cellular
-        let report = make_report(2.0, 348); // rtt_4ms rounds so use 348
+        // RTT 150ms: Degraded on WiFi (>=100ms), Good on cellular (<300ms and loss<8%)
+        let report = make_report(2.0, 148);
         assert_eq!(
             Tier::classify_with_context(&report, NetworkContext::WiFi),
-            Tier::Good
+            Tier::Degraded
         );
         assert_eq!(
             Tier::classify_with_context(&report, NetworkContext::CellularLte),
-            Tier::Degraded
+            Tier::Good
+        );
+    }
+
+    #[test]
+    fn cellular_no_studio_tiers() {
+        // Even with perfect network, cellular stays at Good (no studio)
+        let report = make_report(0.0, 10);
+        assert_eq!(
+            Tier::classify_with_context(&report, NetworkContext::CellularLte),
+            Tier::Good
+        );
+        assert_eq!(
+            Tier::classify_with_context(&report, NetworkContext::WiFi),
+            Tier::Studio64k
         );
     }
 
@@ -469,6 +672,9 @@ mod tests {
 
     #[test]
     fn tier_downgrade() {
+        assert_eq!(Tier::Studio64k.downgrade(), Some(Tier::Studio48k));
+        assert_eq!(Tier::Studio48k.downgrade(), Some(Tier::Studio32k));
+        assert_eq!(Tier::Studio32k.downgrade(), Some(Tier::Good));
         assert_eq!(Tier::Good.downgrade(), Some(Tier::Degraded));
         assert_eq!(Tier::Degraded.downgrade(), Some(Tier::Catastrophic));
         assert_eq!(Tier::Catastrophic.downgrade(), None);
@@ -478,4 +684,97 @@ mod tests {
     fn network_context_default() {
         assert_eq!(NetworkContext::default(), NetworkContext::Unknown);
     }
+
+    // ---------------------------------------------------------------
+    // Bandwidth probing tests
+    // ---------------------------------------------------------------
+
+    #[test]
+    fn probe_triggers_after_stable_period() {
+        let mut ctrl = AdaptiveQualityController::new();
+        let excellent = make_report(0.3, 20); // would classify as Studio64k
+
+        // Starts at Good. Fast-forward stability by setting stable_since directly.
+        ctrl.stable_since = Some(Instant::now() - Duration::from_secs(31));
+
+        // One excellent report should trigger a probe (Good → Studio32k)
+        let result = ctrl.observe(&excellent);
+        assert!(result.is_some(), "should start probe after 30s stable");
+        assert!(ctrl.probe.is_some(), "probe should be active");
+        assert_eq!(ctrl.probe.as_ref().unwrap().target_tier, Tier::Studio32k);
+    }
+
+    #[test]
+    fn probe_succeeds_after_window() {
+        let mut ctrl = AdaptiveQualityController::new();
+        ctrl.stable_since = Some(Instant::now() - Duration::from_secs(31));
+
+        let excellent = make_report(0.3, 20);
+
+        // Trigger probe start
+        let result = ctrl.observe(&excellent);
+        assert!(result.is_some());
+
+        // Simulate probe window elapsed by backdating started
+        ctrl.probe.as_mut().unwrap().started =
+            Instant::now() - Duration::from_secs(PROBE_DURATION_SECS);
+
+        // Next good report should finalize the probe
+        let result = ctrl.observe(&excellent);
+        assert!(result.is_some(), "probe should succeed");
+        assert_eq!(ctrl.current_tier, Tier::Studio32k);
+        assert!(ctrl.probe.is_none(), "probe should be cleared");
+    }
+
+    #[test]
+    fn probe_fails_on_bad_reports() {
+        let mut ctrl = AdaptiveQualityController::new();
+        // Put controller at Studio32k, pretend we've been stable
+        ctrl.current_tier = Tier::Studio32k;
+        ctrl.current_profile = Tier::Studio32k.profile();
+        ctrl.stable_since = Some(Instant::now() - Duration::from_secs(31));
+
+        // Start a probe to Studio48k
+        let excellent = make_report(0.3, 20);
+        let result = ctrl.observe(&excellent);
+        assert!(result.is_some()); // probe started
+        assert_eq!(ctrl.probe.as_ref().unwrap().target_tier, Tier::Studio48k);
+
+        // Feed bad reports (loss too high for Studio48k)
+        let degraded = make_report(3.0, 100);
+        ctrl.observe(&degraded); // first bad
+        ctrl.observe(&degraded); // second bad — exceeds PROBE_MAX_BAD (1)
+
+        // Probe should be cancelled
+        assert!(ctrl.probe.is_none(), "probe should be cancelled after bad reports");
+        // Should still be at Studio32k (not upgraded)
+        assert_eq!(ctrl.current_tier, Tier::Studio32k);
+    }
+
+    #[test]
+    fn no_probe_on_cellular() {
+        let mut ctrl = AdaptiveQualityController::new();
+        ctrl.signal_network_change(NetworkContext::CellularLte);
+        ctrl.current_tier = Tier::Good;
+        ctrl.current_profile = Tier::Good.profile();
+        ctrl.stable_since = Some(Instant::now() - Duration::from_secs(60));
+
+        let good = make_report(0.5, 40);
+        let result = ctrl.observe(&good);
+        // Should NOT probe on cellular
+        assert!(ctrl.probe.is_none(), "should not probe on cellular");
+        assert!(result.is_none() || ctrl.current_tier == Tier::Good);
+    }
+
+    #[test]
+    fn no_probe_at_highest_tier() {
+        let mut ctrl = AdaptiveQualityController::new();
+        ctrl.current_tier = Tier::Studio64k;
+        ctrl.current_profile = Tier::Studio64k.profile();
+        ctrl.stable_since = Some(Instant::now() - Duration::from_secs(60));
+
+        let excellent = make_report(0.1, 10);
+        let result = ctrl.observe(&excellent);
+        assert!(result.is_none(), "should not probe when already at Studio64k");
+    }
 }

crates/wzp-proto/src/traits.rs

@@ -28,6 +28,13 @@ pub trait AudioEncoder: Send + Sync {
 
     /// Enable/disable DTX (discontinuous transmission). No-op for Codec2.
     fn set_dtx(&mut self, _enabled: bool) {}
+
+    /// Hint the encoder about expected packet loss (0–100). In DRED mode the
+    /// encoder floors this at 15% internally. No-op for Codec2.
+    fn set_expected_loss(&mut self, _loss_pct: u8) {}
+
+    /// Set DRED duration in 10 ms frame units (0–104). No-op for Codec2.
+    fn set_dred_duration(&mut self, _frames: u8) {}
 }
 
 /// Decodes compressed frames back to PCM audio.

crates/wzp-relay/Cargo.toml

@@ -20,6 +20,7 @@ bytes = { workspace = true }
 serde = { workspace = true }
 toml = "0.8"
 anyhow = "1"
+clap = { version = "4", features = ["derive"] }
 reqwest = { version = "0.12", features = ["json"] }
 serde_json = "1"
 rustls = { version = "0.23", default-features = false, features = ["ring", "std"] }
@@ -28,6 +29,7 @@ prometheus = "0.13"
 axum = { version = "0.7", default-features = false, features = ["tokio", "http1", "ws"] }
 tower-http = { version = "0.6", features = ["fs"] }
 futures-util = "0.3"
+dashmap = "6"
 dirs = "6"
 sha2 = { workspace = true }
 chrono = "0.4"

crates/wzp-relay/src/call_registry.rs

@@ -31,6 +31,43 @@ pub struct DirectCall {
     pub created_at: Instant,
     pub answered_at: Option<Instant>,
     pub ended_at: Option<Instant>,
+    /// Phase 3 (hole-punching): caller's server-reflexive address
+    /// as carried in the `DirectCallOffer`. The relay stashes it
+    /// here when the offer arrives so it can later inject it as
+    /// `peer_direct_addr` into the callee's `CallSetup`.
+    pub caller_reflexive_addr: Option<String>,
+    /// Phase 3 (hole-punching): callee's server-reflexive address
+    /// as carried in the `DirectCallAnswer`. Only populated for
+    /// `AcceptTrusted` answers — privacy-mode answers leave this
+    /// `None`. Fed into the caller's `CallSetup.peer_direct_addr`.
+    pub callee_reflexive_addr: Option<String>,
+    /// Phase 4 (cross-relay): federation TLS fingerprint of the
+    /// PEER RELAY that forwarded the offer/answer for this call.
+    /// `None` for local calls — caller and callee both
+    /// registered on this relay. `Some(fp)` when one side of
+    /// the call is on a remote relay reached through the
+    /// federation link identified by `fp`. The
+    /// `DirectCallAnswer` handling uses this to route the reply
+    /// back through the SAME link instead of broadcasting again.
+    pub peer_relay_fp: Option<String>,
+    /// Phase 5.5 (ICE host candidates): caller's LAN-local
+    /// interface addresses from the `DirectCallOffer`. Cross-
+    /// wired into the callee's `CallSetup.peer_local_addrs` so
+    /// the callee can direct-dial the caller over the same LAN
+    /// without going through the WAN reflex addr (NAT
+    /// hairpinning often doesn't work for same-LAN peers).
+    pub caller_local_addrs: Vec<String>,
+    /// Phase 5.5 (ICE host candidates): callee's LAN-local
+    /// interface addresses from the `DirectCallAnswer`. Cross-
+    /// wired into the caller's `CallSetup.peer_local_addrs`.
+    pub callee_local_addrs: Vec<String>,
+    /// Phase 8 (Tailscale-inspired): caller's port-mapped
+    /// external address from NAT-PMP/PCP/UPnP. Cross-wired
+    /// into callee's `CallSetup.peer_mapped_addr`.
+    pub caller_mapped_addr: Option<String>,
+    /// Phase 8: callee's port-mapped external address.
+    /// Cross-wired into caller's `CallSetup.peer_mapped_addr`.
+    pub callee_mapped_addr: Option<String>,
 }
 
 /// Registry of active direct calls.
@@ -57,11 +94,79 @@ impl CallRegistry {
             created_at: Instant::now(),
             answered_at: None,
             ended_at: None,
+            caller_reflexive_addr: None,
+            callee_reflexive_addr: None,
+            peer_relay_fp: None,
+            caller_local_addrs: Vec::new(),
+            callee_local_addrs: Vec::new(),
+            caller_mapped_addr: None,
+            callee_mapped_addr: None,
         };
         self.calls.insert(call_id.clone(), call);
         self.calls.get(&call_id).unwrap()
     }
 
+    /// Phase 5.5: stash the caller's LAN host candidates from
+    /// the `DirectCallOffer`. Empty Vec is a valid value meaning
+    /// "caller has no LAN candidates" (e.g. old client).
+    pub fn set_caller_local_addrs(&mut self, call_id: &str, addrs: Vec<String>) {
+        if let Some(call) = self.calls.get_mut(call_id) {
+            call.caller_local_addrs = addrs;
+        }
+    }
+
+    /// Phase 5.5: stash the callee's LAN host candidates from
+    /// the `DirectCallAnswer`.
+    pub fn set_callee_local_addrs(&mut self, call_id: &str, addrs: Vec<String>) {
+        if let Some(call) = self.calls.get_mut(call_id) {
+            call.callee_local_addrs = addrs;
+        }
+    }
+
+    /// Phase 4: stash the federation TLS fingerprint of the peer
+    /// relay that originated (or will receive) the cross-relay
+    /// forward for this call. Safe to call with `None` to clear
+    /// a previously-set value.
+    pub fn set_peer_relay_fp(&mut self, call_id: &str, fp: Option<String>) {
+        if let Some(call) = self.calls.get_mut(call_id) {
+            call.peer_relay_fp = fp;
+        }
+    }
+
+    /// Phase 3: stash the caller's server-reflexive address read
+    /// off a `DirectCallOffer`. Safe to call on any call state;
+    /// a no-op if the call doesn't exist.
+    pub fn set_caller_reflexive_addr(&mut self, call_id: &str, addr: Option<String>) {
+        if let Some(call) = self.calls.get_mut(call_id) {
+            call.caller_reflexive_addr = addr;
+        }
+    }
+
+    /// Phase 3: stash the callee's server-reflexive address read
+    /// off a `DirectCallAnswer`. Safe to call on any call state;
+    /// a no-op if the call doesn't exist.
+    pub fn set_callee_reflexive_addr(&mut self, call_id: &str, addr: Option<String>) {
+        if let Some(call) = self.calls.get_mut(call_id) {
+            call.callee_reflexive_addr = addr;
+        }
+    }
+
+    /// Phase 8: stash the caller's port-mapped address from
+    /// the `DirectCallOffer`.
+    pub fn set_caller_mapped_addr(&mut self, call_id: &str, addr: Option<String>) {
+        if let Some(call) = self.calls.get_mut(call_id) {
+            call.caller_mapped_addr = addr;
+        }
+    }
+
+    /// Phase 8: stash the callee's port-mapped address from
+    /// the `DirectCallAnswer`.
+    pub fn set_callee_mapped_addr(&mut self, call_id: &str, addr: Option<String>) {
+        if let Some(call) = self.calls.get_mut(call_id) {
+            call.callee_mapped_addr = addr;
+        }
+    }
+
     /// Get a call by ID.
     pub fn get(&self, call_id: &str) -> Option<&DirectCall> {
         self.calls.get(call_id)
@@ -196,4 +301,122 @@ mod tests {
         assert_eq!(reg.peer_fingerprint("c1", "alice"), Some("bob"));
         assert_eq!(reg.peer_fingerprint("c1", "bob"), Some("alice"));
     }
+
+    #[test]
+    fn call_registry_stores_reflexive_addrs() {
+        let mut reg = CallRegistry::new();
+        reg.create_call("c1".into(), "alice".into(), "bob".into());
+
+        // Default: both addrs are None.
+        let c = reg.get("c1").unwrap();
+        assert!(c.caller_reflexive_addr.is_none());
+        assert!(c.callee_reflexive_addr.is_none());
+
+        // Caller advertises its reflex addr via DirectCallOffer.
+        reg.set_caller_reflexive_addr("c1", Some("192.0.2.1:4433".into()));
+        assert_eq!(
+            reg.get("c1").unwrap().caller_reflexive_addr.as_deref(),
+            Some("192.0.2.1:4433")
+        );
+
+        // Callee responds with AcceptTrusted + its own reflex addr.
+        reg.set_callee_reflexive_addr("c1", Some("198.51.100.9:4433".into()));
+        assert_eq!(
+            reg.get("c1").unwrap().callee_reflexive_addr.as_deref(),
+            Some("198.51.100.9:4433")
+        );
+
+        // Both addrs are independently readable — the relay uses
+        // them to cross-wire peer_direct_addr in CallSetup.
+        let c = reg.get("c1").unwrap();
+        assert_eq!(
+            c.caller_reflexive_addr.as_deref(),
+            Some("192.0.2.1:4433")
+        );
+        assert_eq!(
+            c.callee_reflexive_addr.as_deref(),
+            Some("198.51.100.9:4433")
+        );
+
+        // Setter on an unknown call is a no-op, not a panic.
+        reg.set_caller_reflexive_addr("does-not-exist", Some("x".into()));
+    }
+
+    #[test]
+    fn call_registry_stores_peer_relay_fp() {
+        let mut reg = CallRegistry::new();
+        reg.create_call("c1".into(), "alice".into(), "bob".into());
+
+        // Default: no peer relay.
+        assert!(reg.get("c1").unwrap().peer_relay_fp.is_none());
+
+        // Cross-relay call: origin relay's fp is stashed.
+        reg.set_peer_relay_fp("c1", Some("relay-a-tls-fp".into()));
+        assert_eq!(
+            reg.get("c1").unwrap().peer_relay_fp.as_deref(),
+            Some("relay-a-tls-fp")
+        );
+
+        // Clearing with None is a valid no-op and empties the field.
+        reg.set_peer_relay_fp("c1", None);
+        assert!(reg.get("c1").unwrap().peer_relay_fp.is_none());
+
+        // Unknown call is a no-op, not a panic.
+        reg.set_peer_relay_fp("does-not-exist", Some("x".into()));
+    }
+
+    #[test]
+    fn call_registry_stores_mapped_addrs() {
+        let mut reg = CallRegistry::new();
+        reg.create_call("c1".into(), "alice".into(), "bob".into());
+
+        // Default: both mapped addrs are None.
+        let c = reg.get("c1").unwrap();
+        assert!(c.caller_mapped_addr.is_none());
+        assert!(c.callee_mapped_addr.is_none());
+
+        // Caller advertises its port-mapped addr via DirectCallOffer.
+        reg.set_caller_mapped_addr("c1", Some("203.0.113.5:12345".into()));
+        assert_eq!(
+            reg.get("c1").unwrap().caller_mapped_addr.as_deref(),
+            Some("203.0.113.5:12345")
+        );
+
+        // Callee responds with its mapped addr.
+        reg.set_callee_mapped_addr("c1", Some("198.51.100.9:54321".into()));
+        assert_eq!(
+            reg.get("c1").unwrap().callee_mapped_addr.as_deref(),
+            Some("198.51.100.9:54321")
+        );
+
+        // Both addrs readable — relay uses them to cross-wire
+        // peer_mapped_addr in CallSetup.
+        let c = reg.get("c1").unwrap();
+        assert_eq!(c.caller_mapped_addr.as_deref(), Some("203.0.113.5:12345"));
+        assert_eq!(c.callee_mapped_addr.as_deref(), Some("198.51.100.9:54321"));
+
+        // Setter on unknown call is a no-op.
+        reg.set_caller_mapped_addr("nope", Some("x".into()));
+    }
+
+    #[test]
+    fn call_registry_clearing_mapped_addr_works() {
+        let mut reg = CallRegistry::new();
+        reg.create_call("c1".into(), "alice".into(), "bob".into());
+        reg.set_caller_mapped_addr("c1", Some("1.2.3.4:5".into()));
+        reg.set_caller_mapped_addr("c1", None);
+        assert!(reg.get("c1").unwrap().caller_mapped_addr.is_none());
+    }
+
+    #[test]
+    fn call_registry_clearing_reflex_addr_works() {
+        // Passing None to the setter must clear a previously-set value
+        // so callers that downgrade to privacy mode mid-flow don't
+        // leak a stale addr into CallSetup.
+        let mut reg = CallRegistry::new();
+        reg.create_call("c1".into(), "alice".into(), "bob".into());
+        reg.set_caller_reflexive_addr("c1", Some("192.0.2.1:4433".into()));
+        reg.set_caller_reflexive_addr("c1", None);
+        assert!(reg.get("c1").unwrap().caller_reflexive_addr.is_none());
+    }
 }

crates/wzp-relay/src/config.rs

@@ -87,6 +87,14 @@ pub struct RelayConfig {
     /// Unlike [[peers]], no url is needed — the peer connects to us.
     #[serde(default)]
     pub trusted: Vec<TrustedConfig>,
+    /// Phase 8: geographic region identifier (e.g., "us-east", "eu-west").
+    /// Sent to clients in `RegisterPresenceAck.relay_region` so they can
+    /// build a relay map for automatic selection.
+    pub region: Option<String>,
+    /// Phase 8: externally-advertised address for this relay. Used to
+    /// populate `available_relays` in `RegisterPresenceAck`. If not set,
+    /// `listen_addr` is used.
+    pub advertised_addr: Option<SocketAddr>,
     /// Debug tap: log packet headers for matching rooms ("*" = all rooms).
     /// Activated via --debug-tap <room> or debug_tap = "room" in TOML.
     pub debug_tap: Option<String>,
@@ -114,6 +122,8 @@ impl Default for RelayConfig {
             peers: Vec::new(),
             global_rooms: Vec::new(),
             trusted: Vec::new(),
+            region: None,
+            advertised_addr: None,
             debug_tap: None,
             event_log: None,
         }

crates/wzp-relay/src/event_log.rs

@@ -5,7 +5,6 @@
 //! Use `wzp-analyzer` to correlate events across multiple relays.
 
 use std::path::PathBuf;
-use std::sync::Arc;
 
 use serde::Serialize;
 use tokio::sync::mpsc;

crates/wzp-relay/src/federation.rs

@@ -134,7 +134,7 @@ pub struct FederationManager {
     peers: Vec<PeerConfig>,
     trusted: Vec<TrustedConfig>,
     global_rooms: HashSet<String>,
-    room_mgr: Arc<Mutex<RoomManager>>,
+    room_mgr: Arc<RoomManager>,
     endpoint: quinn::Endpoint,
     local_tls_fp: String,
     metrics: Arc<crate::metrics::RelayMetrics>,
@@ -142,13 +142,18 @@ pub struct FederationManager {
     peer_links: Arc<Mutex<HashMap<String, PeerLink>>>,
     /// Dedup filter for incoming federation datagrams.
     dedup: Mutex<Deduplicator>,
-    /// Per-room seq counter for federation media delivered to local clients.
-    /// Ensures clients see monotonically increasing seq regardless of federation sender.
-    local_delivery_seq: std::sync::atomic::AtomicU16,
     /// JSONL event log for protocol analysis.
     event_log: EventLogger,
     /// Per-room rate limiters for inbound federation media.
     rate_limiters: Mutex<HashMap<String, RateLimiter>>,
+    /// Phase 4: channel for handing cross-relay direct-call
+    /// signaling (inner message + origin relay fp) back to the
+    /// main signal loop in `main.rs`. Set once at startup via
+    /// `set_cross_relay_tx`. `None` when the main loop hasn't
+    /// wired it up yet (e.g. during startup warmup) — forwards
+    /// that arrive before wiring are dropped with a warning.
+    cross_relay_signal_tx:
+        Mutex<Option<tokio::sync::mpsc::Sender<(wzp_proto::SignalMessage, String)>>>,
 }
 
 impl FederationManager {
@@ -156,7 +161,7 @@ impl FederationManager {
         peers: Vec<PeerConfig>,
         trusted: Vec<TrustedConfig>,
         global_rooms: HashSet<String>,
-        room_mgr: Arc<Mutex<RoomManager>>,
+        room_mgr: Arc<RoomManager>,
         endpoint: quinn::Endpoint,
         local_tls_fp: String,
         metrics: Arc<crate::metrics::RelayMetrics>,
@@ -172,34 +177,138 @@ impl FederationManager {
             metrics,
             peer_links: Arc::new(Mutex::new(HashMap::new())),
             dedup: Mutex::new(Deduplicator::new(DEDUP_WINDOW_SIZE)),
-            local_delivery_seq: std::sync::atomic::AtomicU16::new(0),
             event_log,
             rate_limiters: Mutex::new(HashMap::new()),
+            cross_relay_signal_tx: Mutex::new(None),
+        }
+    }
+
+    /// Phase 4: expose this relay's federation TLS fingerprint so
+    /// the main signal loop can populate
+    /// `SignalMessage::FederatedSignalForward.origin_relay_fp`.
+    pub fn local_tls_fp(&self) -> &str {
+        &self.local_tls_fp
+    }
+
+    /// Phase 4: wire the channel that the main signal loop uses
+    /// to receive unwrapped cross-relay direct-call signals. Called
+    /// once at startup from `main.rs`.
+    pub async fn set_cross_relay_tx(
+        &self,
+        tx: tokio::sync::mpsc::Sender<(wzp_proto::SignalMessage, String)>,
+    ) {
+        *self.cross_relay_signal_tx.lock().await = Some(tx);
+    }
+
+    /// Phase 4: broadcast a `SignalMessage::FederatedSignalForward`
+    /// to every active federation peer link. Returns the number of
+    /// peers the broadcast reached (not the number that successfully
+    /// delivered the message further). Used when the local relay
+    /// doesn't know which peer holds the target fingerprint for a
+    /// `DirectCallOffer` — whichever peer has it will unwrap and
+    /// handle locally; the rest drop silently after "target not
+    /// local" check.
+    ///
+    /// Loop prevention: the receiving relay checks
+    /// `origin_relay_fp` against its own fp and drops self-sourced
+    /// forwards.
+    pub async fn broadcast_signal(&self, msg: &wzp_proto::SignalMessage) -> usize {
+        let peers: Vec<(String, String, Arc<QuinnTransport>)> = {
+            let links = self.peer_links.lock().await;
+            links.iter().map(|(fp, l)| (fp.clone(), l.label.clone(), l.transport.clone())).collect()
+        };  // lock released
+        let mut count = 0;
+        for (fp, label, transport) in &peers {
+            match transport.send_signal(msg).await {
+                Ok(()) => {
+                    count += 1;
+                    tracing::debug!(peer = %label, %fp, "federation: broadcast signal ok");
+                }
+                Err(e) => {
+                    tracing::warn!(peer = %label, %fp, error = %e, "federation: broadcast signal failed");
+                }
+            }
+        }
+        count
+    }
+
+    /// Phase 4: targeted send — used by the
+    /// `DirectCallAnswer` path when the registry knows exactly
+    /// which peer relay to route the reply back to. More efficient
+    /// than re-broadcasting and avoids leaking the call to
+    /// uninvolved peers.
+    ///
+    /// Returns `Ok(())` on success, `Err(String)` when the peer
+    /// isn't currently linked or the send fails.
+    pub async fn send_signal_to_peer(
+        &self,
+        peer_relay_fp: &str,
+        msg: &wzp_proto::SignalMessage,
+    ) -> Result<(), String> {
+        let normalized = normalize_fp(peer_relay_fp);
+        let transport = {
+            let links = self.peer_links.lock().await;
+            links.get(&normalized).map(|l| l.transport.clone())
+        };  // lock released
+        match transport {
+            Some(t) => t
+                .send_signal(msg)
+                .await
+                .map_err(|e| format!("send to peer {normalized}: {e}")),
+            None => Err(format!("no active federation link for {normalized}")),
         }
     }
 
     /// Check if a room name (which may be hashed) is a global room.
+    ///
+    /// Phase 4.1: ALL `call-*` rooms are implicitly global for
+    /// federation. This is the simplest path to cross-relay direct
+    /// calling with relay-mediated media fallback: when both peers
+    /// join the same `call-<id>` room on their respective relays,
+    /// the federation media pipeline automatically forwards
+    /// datagrams between them. The relay's existing ACL (`call-*`
+    /// rooms are restricted to the two authorized participants in
+    /// the call registry) prevents random clients from creating or
+    /// joining `call-*` rooms.
     pub fn is_global_room(&self, room: &str) -> bool {
+        if room.starts_with("call-") {
+            return true;
+        }
         self.resolve_global_room(room).is_some()
     }
 
     /// Resolve a room name (raw or hashed) to the canonical global room name.
     /// Returns the configured global room name if it matches.
-    pub fn resolve_global_room(&self, room: &str) -> Option<&str> {
+    ///
+    /// Phase 4.1: `call-*` rooms resolve to themselves (they ARE
+    /// the canonical name — no hashing or aliasing involved).
+    ///
+    /// Returns `Option<String>` (owned) instead of `Option<&str>`
+    /// because call-* room names aren't stored on `self` — they
+    /// come from the caller and we just confirm "yes, this is
+    /// global" by returning it back. Pre-4.1 callers that used
+    /// the reference for equality checks or hashing work
+    /// unchanged via String/&str auto-deref.
+    pub fn resolve_global_room(&self, room: &str) -> Option<String> {
+        // Phase 4.1: call-* rooms are implicitly global, resolve
+        // to themselves
+        if room.starts_with("call-") {
+            return Some(room.to_string());
+        }
         // Direct match (raw room name, e.g. Android clients)
         if self.global_rooms.contains(room) {
-            return Some(self.global_rooms.iter().find(|n| n.as_str() == room).unwrap());
+            return Some(room.to_string());
         }
         // Hashed match (desktop clients hash room names for SNI privacy)
         self.global_rooms.iter().find(|name| {
             wzp_crypto::hash_room_name(name) == room
-        }).map(|s| s.as_str())
+        }).map(|s| s.to_string())
     }
 
     /// Get the canonical federation room hash for a room.
     /// Always uses the configured global room name, not the client-provided name.
     pub fn global_room_hash(&self, room: &str) -> [u8; 8] {
-        if let Some(canonical) = self.resolve_global_room(room) {
+        if let Some(ref canonical) = self.resolve_global_room(room) {
             room_hash(canonical)
         } else {
             room_hash(room)
@@ -229,10 +338,7 @@ impl FederationManager {
         }
 
         // Room event dispatcher
-        let room_events = {
-            let mgr = self.room_mgr.lock().await;
-            mgr.subscribe_events()
-        };
+        let room_events = self.room_mgr.subscribe_events();
         let this = self.clone();
         handles.push(tokio::spawn(async move {
             run_room_event_dispatcher(this, room_events).await;
@@ -271,8 +377,8 @@ impl FederationManager {
         let mut result = Vec::new();
         for link in links.values() {
             // Check canonical name
-            if let Some(c) = canonical {
-                if let Some(remote) = link.remote_participants.get(c) {
+            if let Some(ref c) = canonical {
+                if let Some(remote) = link.remote_participants.get(c.as_str()) {
                     result.extend(remote.iter().cloned());
                 }
                 // Also check raw room name, but only if different from canonical
@@ -296,21 +402,28 @@ impl FederationManager {
     /// Forward locally-generated media to all connected peers.
     /// For locally-originated media, we send to ALL peers (they decide whether to deliver).
     /// For forwarded media (multi-hop), handle_datagram filters by active_rooms.
-    pub async fn forward_to_peers(&self, room_name: &str, room_hash: &[u8; 8], media_data: &Bytes) {
-        let links = self.peer_links.lock().await;
-        if links.is_empty() {
-            return;
-        }
-        for (_fp, link) in links.iter() {
+    ///
+    /// `_room_name` is kept in the signature for caller-site symmetry with
+    /// the other room-tagged helpers and for future per-room-name logging
+    /// or rate limiting; the body currently forwards on `room_hash` alone
+    /// because that's what the wire format carries.
+    pub async fn forward_to_peers(&self, _room_name: &str, room_hash: &[u8; 8], media_data: &Bytes) {
+        let peers: Vec<(String, Arc<QuinnTransport>)> = {
+            let links = self.peer_links.lock().await;
+            if links.is_empty() { return; }
+            links.values().map(|l| (l.label.clone(), l.transport.clone())).collect()
+        };  // lock released
+
+        for (label, transport) in &peers {
             let mut tagged = Vec::with_capacity(8 + media_data.len());
             tagged.extend_from_slice(room_hash);
             tagged.extend_from_slice(media_data);
-            match link.transport.send_raw_datagram(&tagged) {
+            match transport.send_raw_datagram(&tagged) {
                 Ok(()) => {
                     self.metrics.federation_packets_forwarded
-                        .with_label_values(&[&link.label, "out"]).inc();
+                        .with_label_values(&[label, "out"]).inc();
                 }
-                Err(e) => warn!(peer = %link.label, "federation send error: {e}"),
+                Err(e) => warn!(peer = %label, "federation send error: {e}"),
             }
         }
     }
@@ -374,15 +487,15 @@ async fn run_room_event_dispatcher(
         match events.recv().await {
             Ok(RoomEvent::LocalJoin { room }) => {
                 if fm.is_global_room(&room) {
-                    let participants = {
-                        let mgr = fm.room_mgr.lock().await;
-                        mgr.local_participant_list(&room)
-                    };
+                    let participants = fm.room_mgr.local_participant_list(&room);
                     info!(room = %room, count = participants.len(), "global room now active, announcing to peers");
                     let msg = SignalMessage::GlobalRoomActive { room, participants };
-                    let links = fm.peer_links.lock().await;
-                    for link in links.values() {
-                        let _ = link.transport.send_signal(&msg).await;
+                    let transports: Vec<Arc<QuinnTransport>> = {
+                        let links = fm.peer_links.lock().await;
+                        links.values().map(|l| l.transport.clone()).collect()
+                    };
+                    for t in &transports {
+                        let _ = t.send_signal(&msg).await;
                     }
                 }
             }
@@ -390,9 +503,12 @@ async fn run_room_event_dispatcher(
                 if fm.is_global_room(&room) {
                     info!(room = %room, "global room now inactive, announcing to peers");
                     let msg = SignalMessage::GlobalRoomInactive { room };
-                    let links = fm.peer_links.lock().await;
-                    for link in links.values() {
-                        let _ = link.transport.send_signal(&msg).await;
+                    let transports: Vec<Arc<QuinnTransport>> = {
+                        let links = fm.peer_links.lock().await;
+                        links.values().map(|l| l.transport.clone()).collect()
+                    };
+                    for t in &transports {
+                        let _ = t.send_signal(&msg).await;
                     }
                 }
             }
@@ -451,11 +567,11 @@ async fn run_stale_presence_sweeper(fm: Arc<FederationManager>) {
 
         // Broadcast updated RoomUpdate for affected rooms
         for room in &affected_rooms {
-            let mgr = fm.room_mgr.lock().await;
-            for local_room in mgr.active_rooms() {
-                if fm.resolve_global_room(&local_room) == fm.resolve_global_room(room) {
-                    let mut all_participants = mgr.local_participant_list(&local_room);
-                    let remote = fm.get_remote_participants(&local_room).await;
+            let active = fm.room_mgr.active_rooms();
+            for local_room in &active {
+                if fm.resolve_global_room(local_room) == fm.resolve_global_room(room) {
+                    let mut all_participants = fm.room_mgr.local_participant_list(local_room);
+                    let remote = fm.get_remote_participants(local_room).await;
                     all_participants.extend(remote);
                     let mut seen = HashSet::new();
                     all_participants.retain(|p| seen.insert(p.fingerprint.clone()));
@@ -463,8 +579,7 @@ async fn run_stale_presence_sweeper(fm: Arc<FederationManager>) {
                         count: all_participants.len() as u32,
                         participants: all_participants,
                     };
-                    let senders = mgr.local_senders(&local_room);
-                    drop(mgr);
+                    let senders = fm.room_mgr.local_senders(local_room);
                     room::broadcast_signal(&senders, &update).await;
                     info!(room = %room, "swept stale presence — broadcast updated RoomUpdate");
                     break;
@@ -542,14 +657,13 @@ async fn run_federation_link(
     // Announce our currently active global rooms to this new peer
     // Collect all announcements first, then send (avoid holding locks across await)
     let announcements = {
-        let mgr = fm.room_mgr.lock().await;
-        let active = mgr.active_rooms();
+        let active = fm.room_mgr.active_rooms();
         let mut msgs = Vec::new();
 
         // Local rooms
         for room_name in &active {
             if fm.is_global_room(room_name) {
-                let participants = mgr.local_participant_list(room_name);
+                let participants = fm.room_mgr.local_participant_list(room_name);
                 info!(peer = %peer_label, room = %room_name, participants = participants.len(), "announcing local global room to new peer");
                 msgs.push(SignalMessage::GlobalRoomActive { room: room_name.clone(), participants });
             }
@@ -623,11 +737,20 @@ async fn run_federation_link(
         }
     };
 
-    // RTT monitor: periodically sample QUIC RTT for this peer
+    // RTT monitor: periodically sample QUIC RTT for this peer and push it
+    // into the `wzp_federation_peer_rtt_ms` gauge. The gauge is registered
+    // in metrics.rs but previously never received any samples — the task
+    // computed rtt_ms and dropped it on the floor, leaving the Grafana
+    // panel blank. Fixed as part of the workspace warning sweep.
     let rtt_task = async move {
         loop {
             tokio::time::sleep(Duration::from_secs(5)).await;
             let rtt_ms = rtt_transport.connection().stats().path.rtt.as_millis() as f64;
+            fm_rtt
+                .metrics
+                .federation_peer_rtt_ms
+                .with_label_values(&[&label_rtt])
+                .set(rtt_ms);
         }
     };
 
@@ -710,22 +833,24 @@ async fn handle_signal(
 
                 // Broadcast updated RoomUpdate to local clients in this room
                 // Find the local room name (may be hashed or raw)
-                let mgr = fm.room_mgr.lock().await;
-                for local_room in mgr.active_rooms() {
-                    if fm.is_global_room(&local_room) && fm.resolve_global_room(&local_room) == fm.resolve_global_room(&room) {
+                let active = fm.room_mgr.active_rooms();
+                for local_room in &active {
+                    if fm.is_global_room(local_room) && fm.resolve_global_room(local_room) == fm.resolve_global_room(&room) {
                         // Build merged participant list: local + all remote (deduped)
-                        let mut all_participants = mgr.local_participant_list(&local_room);
-                        let links = fm.peer_links.lock().await;
-                        for link in links.values() {
-                            if let Some(canonical) = fm.resolve_global_room(&local_room) {
-                                if let Some(remote) = link.remote_participants.get(canonical) {
-                                    all_participants.extend(remote.iter().cloned());
-                                }
-                                // Also check raw room name, but only if different from canonical
-                                if canonical != local_room {
-                                    if let Some(remote) = link.remote_participants.get(&local_room) {
+                        let mut all_participants = fm.room_mgr.local_participant_list(local_room);
+                        {
+                            let links = fm.peer_links.lock().await;
+                            for link in links.values() {
+                                if let Some(ref canonical) = fm.resolve_global_room(local_room) {
+                                    if let Some(remote) = link.remote_participants.get(canonical.as_str()) {
                                         all_participants.extend(remote.iter().cloned());
                                     }
+                                    // Also check raw room name, but only if different from canonical
+                                    if canonical != local_room {
+                                        if let Some(remote) = link.remote_participants.get(local_room) {
+                                            all_participants.extend(remote.iter().cloned());
+                                        }
+                                    }
                                 }
                             }
                         }
@@ -736,9 +861,7 @@ async fn handle_signal(
                             count: all_participants.len() as u32,
                             participants: all_participants,
                         };
-                        let senders = mgr.local_senders(&local_room);
-                        drop(links);
-                        drop(mgr);
+                        let senders = fm.room_mgr.local_senders(local_room);
                         room::broadcast_signal(&senders, &update).await;
                         break;
                     }
@@ -753,8 +876,8 @@ async fn handle_signal(
                 // Clear remote participants for this peer+room
                 link.remote_participants.remove(&room);
                 // Also try canonical name
-                if let Some(canonical) = fm.resolve_global_room(&room) {
-                    link.remote_participants.remove(canonical);
+                if let Some(ref canonical) = fm.resolve_global_room(&room) {
+                    link.remote_participants.remove(canonical.as_str());
                 }
             }
 
@@ -768,8 +891,8 @@ async fn handle_signal(
                 let mut result = Vec::new();
                 for (fp, link) in links.iter() {
                     if fp == peer_fp { continue; }
-                    if let Some(c) = canonical {
-                        if let Some(remote) = link.remote_participants.get(c) {
+                    if let Some(ref c) = canonical {
+                        if let Some(remote) = link.remote_participants.get(c.as_str()) {
                             result.extend(remote.iter().cloned());
                         }
                     }
@@ -781,10 +904,7 @@ async fn handle_signal(
 
             // Propagate to other peers: send updated GlobalRoomActive with revised list,
             // or GlobalRoomInactive if no participants remain anywhere
-            let local_active = {
-                let mgr = fm.room_mgr.lock().await;
-                mgr.active_rooms().iter().any(|r| fm.resolve_global_room(r) == fm.resolve_global_room(&room))
-            };
+            let local_active = fm.room_mgr.active_rooms().iter().any(|r| fm.resolve_global_room(r) == fm.resolve_global_room(&room));
             let has_remaining = !remaining_remote.is_empty() || local_active;
 
             // Collect peer transports to send to (avoid holding lock across await)
@@ -798,10 +918,9 @@ async fn handle_signal(
                 // Send updated participant list to other peers
                 let mut updated_participants = remaining_remote.clone();
                 if local_active {
-                    let mgr = fm.room_mgr.lock().await;
-                    for local_room in mgr.active_rooms() {
+                    for local_room in fm.room_mgr.active_rooms() {
                         if fm.resolve_global_room(&local_room) == fm.resolve_global_room(&room) {
-                            updated_participants.extend(mgr.local_participant_list(&local_room));
+                            updated_participants.extend(fm.room_mgr.local_participant_list(&local_room));
                             break;
                         }
                     }
@@ -822,10 +941,10 @@ async fn handle_signal(
             }
 
             // Broadcast updated RoomUpdate to local clients (remote participant removed)
-            let mgr = fm.room_mgr.lock().await;
-            for local_room in mgr.active_rooms() {
-                if fm.is_global_room(&local_room) && fm.resolve_global_room(&local_room) == fm.resolve_global_room(&room) {
-                    let mut all_participants = mgr.local_participant_list(&local_room);
+            let active = fm.room_mgr.active_rooms();
+            for local_room in &active {
+                if fm.is_global_room(local_room) && fm.resolve_global_room(local_room) == fm.resolve_global_room(&room) {
+                    let mut all_participants = fm.room_mgr.local_participant_list(local_room);
                     all_participants.extend(remaining_remote.iter().cloned());
                     // Deduplicate by fingerprint
                     let mut seen = HashSet::new();
@@ -834,14 +953,64 @@ async fn handle_signal(
                         count: all_participants.len() as u32,
                         participants: all_participants,
                     };
-                    let senders = mgr.local_senders(&local_room);
-                    drop(mgr);
+                    let senders = fm.room_mgr.local_senders(local_room);
                     room::broadcast_signal(&senders, &update).await;
                     info!(room = %room, "broadcast updated presence (remote participant removed)");
                     break;
                 }
             }
         }
+        // Phase 4: cross-relay direct-call signal envelope.
+        //
+        // Unwrap the inner message and hand it off to the main
+        // signal loop via the cross_relay_signal_tx channel. The
+        // main loop will then dispatch the inner DirectCallOffer/
+        // Answer/Ringing/Hangup exactly as if it had arrived on a
+        // local signal transport — with the extra context that
+        // the call is "federated" (origin_relay_fp).
+        //
+        // Loop prevention: drop any forward whose origin matches
+        // our own federation TLS fingerprint. With
+        // broadcast-to-all-peers this prevents A→B→A echo loops.
+        SignalMessage::FederatedSignalForward { inner, origin_relay_fp } => {
+            if origin_relay_fp == fm.local_tls_fp {
+                tracing::debug!(
+                    peer = %peer_label,
+                    "federation: dropping self-sourced FederatedSignalForward (loop prevention)"
+                );
+                return;
+            }
+            let tx_opt = {
+                let guard = fm.cross_relay_signal_tx.lock().await;
+                guard.clone()
+            };
+            match tx_opt {
+                Some(tx) => {
+                    let inner_discriminant = std::mem::discriminant(&*inner);
+                    if let Err(e) = tx.send((*inner, origin_relay_fp.clone())).await {
+                        warn!(
+                            peer = %peer_label,
+                            ?inner_discriminant,
+                            error = %e,
+                            "federation: cross-relay signal dispatcher full / closed"
+                        );
+                    } else {
+                        tracing::debug!(
+                            peer = %peer_label,
+                            ?inner_discriminant,
+                            %origin_relay_fp,
+                            "federation: forwarded cross-relay signal to main dispatcher"
+                        );
+                    }
+                }
+                None => {
+                    warn!(
+                        peer = %peer_label,
+                        "federation: cross_relay_signal_tx not wired yet — dropping forward"
+                    );
+                }
+            }
+        }
         _ => {} // ignore other signals
     }
 }
@@ -901,14 +1070,13 @@ async fn handle_datagram(
         }
     }
 
-    // Find room by hash — check local rooms AND global room config
+    // Find room by hash -- check local rooms AND global room config
     let room_name = {
-        let mgr = fm.room_mgr.lock().await;
-        let active = mgr.active_rooms();
+        let active = fm.room_mgr.active_rooms();
         // First: check local rooms (has participants)
         active.iter().find(|r| room_hash(r) == rh).cloned()
             .or_else(|| active.iter().find(|r| fm.global_room_hash(r) == rh).cloned())
-            // Second: check global room config (hub relay may have no local participants)
+            // Second: check static global room config (hub relay may have no local participants)
             .or_else(|| {
                 fm.global_rooms.iter().find(|name| room_hash(name) == rh).cloned()
             })
@@ -918,6 +1086,20 @@ async fn handle_datagram(
         Some(r) => r,
         None => {
             fm.event_log.emit(Event::new("room_not_found").seq(pkt.header.seq).peer(&peer_label));
+            // Phase 4.1 diagnostic: log the hash + active rooms
+            // so we can diagnose cross-relay call-* media routing
+            // failures. This fires when a peer relay sends media
+            // for a room we don't have locally — could be a
+            // timing issue (peer joined before us) or a hash
+            // mismatch.
+            let active = fm.room_mgr.active_rooms();
+            warn!(
+                room_hash = ?rh,
+                active_rooms = ?active,
+                seq = pkt.header.seq,
+                peer = %peer_label,
+                "federation datagram for unknown room — no local room matches hash"
+            );
             return;
         }
     };
@@ -935,10 +1117,7 @@ async fn handle_datagram(
 
     // Deliver to all local participants — forward the raw bytes as-is.
     // The original sender's MediaPacket is preserved exactly (no re-serialization).
-    let locals = {
-        let mgr = fm.room_mgr.lock().await;
-        mgr.local_senders(&room_name)
-    };
+    let locals = fm.room_mgr.local_senders(&room_name);
     for sender in &locals {
         match sender {
             room::ParticipantSender::Quic(t) => {

crates/wzp-relay/src/handshake.rs

@@ -94,9 +94,13 @@ pub async fn accept_handshake(
 }
 
 /// Select the best quality profile from those the caller supports.
-fn choose_profile(supported: &[QualityProfile]) -> QualityProfile {
-    // Cap at GOOD (24k) for now — studio tiers (32k/48k/64k) not yet tested
-    // for federation reliability (large packets may exceed path MTU).
+///
+/// The `_supported` list is currently ignored — we hardcode GOOD (24k) until
+/// studio tiers (32k/48k/64k) have been validated across federation (large
+/// packets may exceed path MTU and fragment in unpleasant ways). Once that's
+/// tested, the body should pick the highest supported profile ≤ the relay's
+/// configured ceiling.
+fn choose_profile(_supported: &[QualityProfile]) -> QualityProfile {
     QualityProfile::GOOD
 }
 

crates/wzp-relay/src/metrics.rs

@@ -29,6 +29,9 @@ pub struct RelayMetrics {
     pub session_rtt_ms: GaugeVec,
     pub session_underruns: IntCounterVec,
     pub session_overruns: IntCounterVec,
+    // Phase 4: loss-recovery breakdown per session.
+    pub session_dred_reconstructions: IntCounterVec,
+    pub session_classical_plc: IntCounterVec,
     registry: Registry,
 }
 
@@ -130,6 +133,23 @@ impl RelayMetrics {
         )
         .expect("metric");
 
+        let session_dred_reconstructions = IntCounterVec::new(
+            Opts::new(
+                "wzp_relay_session_dred_reconstructions_total",
+                "Frames reconstructed via DRED (Deep REDundancy) per session",
+            ),
+            &["session_id"],
+        )
+        .expect("metric");
+        let session_classical_plc = IntCounterVec::new(
+            Opts::new(
+                "wzp_relay_session_classical_plc_total",
+                "Frames filled via classical Opus/Codec2 PLC per session",
+            ),
+            &["session_id"],
+        )
+        .expect("metric");
+
         registry.register(Box::new(active_sessions.clone())).expect("register");
         registry.register(Box::new(active_rooms.clone())).expect("register");
         registry.register(Box::new(packets_forwarded.clone())).expect("register");
@@ -147,6 +167,8 @@ impl RelayMetrics {
         registry.register(Box::new(session_rtt_ms.clone())).expect("register");
         registry.register(Box::new(session_underruns.clone())).expect("register");
         registry.register(Box::new(session_overruns.clone())).expect("register");
+        registry.register(Box::new(session_dred_reconstructions.clone())).expect("register");
+        registry.register(Box::new(session_classical_plc.clone())).expect("register");
 
         Self {
             active_sessions,
@@ -166,6 +188,8 @@ impl RelayMetrics {
             session_rtt_ms,
             session_underruns,
             session_overruns,
+            session_dred_reconstructions,
+            session_classical_plc,
             registry,
         }
     }
@@ -217,6 +241,39 @@ impl RelayMetrics {
         }
     }
 
+    /// Phase 4: update per-session loss-recovery counters from a client's
+    /// `LossRecoveryUpdate` signal message. The client sends monotonic
+    /// totals (frames reconstructed since call start); we compute the
+    /// delta against the current Prometheus counter and increment by it.
+    /// IntCounterVec only increases, so a client restart that resets the
+    /// counter to 0 simply produces no delta until the new totals exceed
+    /// the Prometheus state.
+    pub fn update_session_loss_recovery(
+        &self,
+        session_id: &str,
+        dred_reconstructions: u64,
+        classical_plc: u64,
+    ) {
+        let cur_dred = self
+            .session_dred_reconstructions
+            .with_label_values(&[session_id])
+            .get();
+        if dred_reconstructions > cur_dred {
+            self.session_dred_reconstructions
+                .with_label_values(&[session_id])
+                .inc_by(dred_reconstructions - cur_dred);
+        }
+        let cur_plc = self
+            .session_classical_plc
+            .with_label_values(&[session_id])
+            .get();
+        if classical_plc > cur_plc {
+            self.session_classical_plc
+                .with_label_values(&[session_id])
+                .inc_by(classical_plc - cur_plc);
+        }
+    }
+
     /// Remove all per-session label values for a disconnected session.
     pub fn remove_session_metrics(&self, session_id: &str) {
         let _ = self.session_buffer_depth.remove_label_values(&[session_id]);
@@ -224,6 +281,10 @@ impl RelayMetrics {
         let _ = self.session_rtt_ms.remove_label_values(&[session_id]);
         let _ = self.session_underruns.remove_label_values(&[session_id]);
         let _ = self.session_overruns.remove_label_values(&[session_id]);
+        let _ = self
+            .session_dred_reconstructions
+            .remove_label_values(&[session_id]);
+        let _ = self.session_classical_plc.remove_label_values(&[session_id]);
     }
 
     /// Get a reference to the underlying Prometheus registry.
@@ -418,10 +479,13 @@ mod tests {
         };
         m.update_session_quality("sess-cleanup", &report);
         m.update_session_buffer("sess-cleanup", 42, 3, 1);
+        m.update_session_loss_recovery("sess-cleanup", 17, 4);
 
         // Verify they appear
         let output = m.metrics_handler();
         assert!(output.contains("sess-cleanup"));
+        assert!(output.contains("wzp_relay_session_dred_reconstructions_total"));
+        assert!(output.contains("wzp_relay_session_classical_plc_total"));
 
         // Remove and verify they are gone
         m.remove_session_metrics("sess-cleanup");
@@ -429,6 +493,55 @@ mod tests {
         assert!(!output.contains("sess-cleanup"));
     }
 
+    /// Phase 4: LossRecoveryUpdate → per-session counters, monotonic delta
+    /// application.
+    #[test]
+    fn session_loss_recovery_monotonic_delta() {
+        let m = RelayMetrics::new();
+        let sess = "sess-dred";
+
+        // First update: 10 DRED, 2 PLC
+        m.update_session_loss_recovery(sess, 10, 2);
+        let dred1 = m
+            .session_dred_reconstructions
+            .with_label_values(&[sess])
+            .get();
+        let plc1 = m.session_classical_plc.with_label_values(&[sess]).get();
+        assert_eq!(dred1, 10);
+        assert_eq!(plc1, 2);
+
+        // Second update: 25 DRED, 5 PLC — counter advances by (15, 3)
+        m.update_session_loss_recovery(sess, 25, 5);
+        let dred2 = m
+            .session_dred_reconstructions
+            .with_label_values(&[sess])
+            .get();
+        let plc2 = m.session_classical_plc.with_label_values(&[sess]).get();
+        assert_eq!(dred2, 25);
+        assert_eq!(plc2, 5);
+
+        // Third update with LOWER values (e.g., client reset) — counters
+        // hold steady, no decrement.
+        m.update_session_loss_recovery(sess, 5, 1);
+        let dred3 = m
+            .session_dred_reconstructions
+            .with_label_values(&[sess])
+            .get();
+        let plc3 = m.session_classical_plc.with_label_values(&[sess]).get();
+        assert_eq!(dred3, 25, "counter must not decrease");
+        assert_eq!(plc3, 5, "counter must not decrease");
+
+        // Fourth update: client caught up and exceeded the old max.
+        m.update_session_loss_recovery(sess, 30, 8);
+        let dred4 = m
+            .session_dred_reconstructions
+            .with_label_values(&[sess])
+            .get();
+        let plc4 = m.session_classical_plc.with_label_values(&[sess]).get();
+        assert_eq!(dred4, 30);
+        assert_eq!(plc4, 8);
+    }
+
     #[test]
     fn metrics_increment() {
         let m = RelayMetrics::new();

crates/wzp-relay/src/room.rs

@@ -9,10 +9,12 @@ use std::sync::Arc;
 use std::time::Duration;
 
 use bytes::Bytes;
-use tokio::sync::Mutex;
-use tracing::{debug, error, info, trace, warn};
+use dashmap::DashMap;
+use tracing::{error, info, warn};
 
 use wzp_proto::packet::TrunkFrame;
+use wzp_proto::quality::{AdaptiveQualityController, Tier};
+use wzp_proto::traits::QualityController;
 use wzp_proto::MediaTransport;
 
 use crate::metrics::RelayMetrics;
@@ -48,6 +50,143 @@ impl DebugTap {
             "TAP"
         );
     }
+
+    pub fn log_signal(&self, room: &str, signal: &wzp_proto::SignalMessage) {
+        match signal {
+            wzp_proto::SignalMessage::RoomUpdate { count, participants } => {
+                let names: Vec<&str> = participants.iter()
+                    .map(|p| p.alias.as_deref().unwrap_or("?"))
+                    .collect();
+                info!(
+                    target: "debug_tap",
+                    room = %room,
+                    signal = "RoomUpdate",
+                    count,
+                    participants = ?names,
+                    "TAP SIGNAL"
+                );
+            }
+            wzp_proto::SignalMessage::QualityDirective { recommended_profile, reason } => {
+                info!(
+                    target: "debug_tap",
+                    room = %room,
+                    signal = "QualityDirective",
+                    codec = ?recommended_profile.codec,
+                    reason = reason.as_deref().unwrap_or(""),
+                    "TAP SIGNAL"
+                );
+            }
+            other => {
+                info!(
+                    target: "debug_tap",
+                    room = %room,
+                    signal = ?std::mem::discriminant(other),
+                    "TAP SIGNAL"
+                );
+            }
+        }
+    }
+
+    pub fn log_event(&self, room: &str, event: &str, detail: &str) {
+        info!(
+            target: "debug_tap",
+            room = %room,
+            event,
+            detail,
+            "TAP EVENT"
+        );
+    }
+
+    pub fn log_stats(&self, room: &str, stats: &TapStats) {
+        let codecs: Vec<String> = stats.codecs_seen.iter().map(|c| format!("{c:?}")).collect();
+        info!(
+            target: "debug_tap",
+            room = %room,
+            period = "5s",
+            in_pkts = stats.in_pkts,
+            out_pkts = stats.out_pkts,
+            fan_out_avg = format!("{:.1}", if stats.in_pkts > 0 { stats.out_pkts as f64 / stats.in_pkts as f64 } else { 0.0 }),
+            seq_gaps = stats.seq_gaps,
+            codecs_seen = ?codecs,
+            "TAP STATS"
+        );
+    }
+}
+
+/// Per-participant stats for the debug tap periodic summary.
+pub struct TapStats {
+    pub in_pkts: u64,
+    pub out_pkts: u64,
+    pub seq_gaps: u64,
+    pub codecs_seen: std::collections::HashSet<wzp_proto::CodecId>,
+    last_seq: Option<u16>,
+}
+
+impl TapStats {
+    pub fn new() -> Self {
+        Self {
+            in_pkts: 0,
+            out_pkts: 0,
+            seq_gaps: 0,
+            codecs_seen: std::collections::HashSet::new(),
+            last_seq: None,
+        }
+    }
+
+    pub fn record_in(&mut self, pkt: &wzp_proto::MediaPacket, fan_out: usize) {
+        self.in_pkts += 1;
+        self.out_pkts += fan_out as u64;
+        self.codecs_seen.insert(pkt.header.codec_id);
+        if let Some(prev) = self.last_seq {
+            let expected = prev.wrapping_add(1);
+            if pkt.header.seq != expected {
+                self.seq_gaps += 1;
+            }
+        }
+        self.last_seq = Some(pkt.header.seq);
+    }
+
+    pub fn reset_period(&mut self) {
+        self.in_pkts = 0;
+        self.out_pkts = 0;
+        self.seq_gaps = 0;
+        // Keep codecs_seen and last_seq across periods
+    }
+}
+
+/// Tracks network quality for a single participant in a room.
+struct ParticipantQuality {
+    controller: AdaptiveQualityController,
+    current_tier: Tier,
+}
+
+impl ParticipantQuality {
+    fn new() -> Self {
+        Self {
+            controller: AdaptiveQualityController::new(),
+            current_tier: Tier::Good,
+        }
+    }
+
+    /// Feed a quality report and return the new tier if it changed.
+    fn observe(&mut self, report: &wzp_proto::packet::QualityReport) -> Option<Tier> {
+        let _ = self.controller.observe(report);
+        let new_tier = self.controller.tier();
+        if new_tier != self.current_tier {
+            self.current_tier = new_tier;
+            Some(new_tier)
+        } else {
+            None
+        }
+    }
+}
+
+/// Compute the weakest (worst) quality tier across all tracked participants.
+fn weakest_tier<'a>(qualities: impl Iterator<Item = &'a ParticipantQuality>) -> Tier {
+    qualities
+        .map(|pq| pq.current_tier)
+        .min()
+        .unwrap_or(Tier::Good)
 }
 
 /// Unique participant ID within a room.
@@ -138,12 +277,18 @@ struct Participant {
 /// A room holding multiple participants.
 struct Room {
     participants: Vec<Participant>,
+    /// Per-participant quality tracking, keyed by participant_id.
+    qualities: HashMap<ParticipantId, ParticipantQuality>,
+    /// Current room-wide tier (to avoid repeated broadcasts).
+    current_tier: Tier,
 }
 
 impl Room {
     fn new() -> Self {
         Self {
             participants: Vec::new(),
+            qualities: HashMap::new(),
+            current_tier: Tier::Good,
         }
     }
 
@@ -200,12 +345,16 @@ impl Room {
 }
 
 /// Manages all rooms on the relay.
+///
+/// Uses `DashMap` for per-room sharded locking -- rooms are independently
+/// lockable so the media hot-path never contends on a single mutex.
 pub struct RoomManager {
-    rooms: HashMap<String, Room>,
-    /// Room access control list. Maps hashed room name → allowed fingerprints.
+    rooms: DashMap<String, Room>,
+    /// Room access control list. Maps hashed room name -> allowed fingerprints.
     /// When `None`, rooms are open (no auth mode). When `Some`, only listed
-    /// fingerprints can join the corresponding room.
-    acl: Option<HashMap<String, HashSet<String>>>,
+    /// fingerprints can join the corresponding room. Protected by std Mutex
+    /// since ACL mutations are rare (only during call setup).
+    acl: Option<std::sync::Mutex<HashMap<String, HashSet<String>>>>,
     /// Channel for room lifecycle events (federation subscribes).
     event_tx: tokio::sync::broadcast::Sender<RoomEvent>,
 }
@@ -214,7 +363,7 @@ impl RoomManager {
     pub fn new() -> Self {
         let (event_tx, _) = tokio::sync::broadcast::channel(64);
         Self {
-            rooms: HashMap::new(),
+            rooms: DashMap::new(),
             acl: None,
             event_tx,
         }
@@ -224,8 +373,8 @@ impl RoomManager {
     pub fn with_acl() -> Self {
         let (event_tx, _) = tokio::sync::broadcast::channel(64);
         Self {
-            rooms: HashMap::new(),
-            acl: Some(HashMap::new()),
+            rooms: DashMap::new(),
+            acl: Some(std::sync::Mutex::new(HashMap::new())),
             event_tx,
         }
     }
@@ -236,9 +385,10 @@ impl RoomManager {
     }
 
     /// Grant a fingerprint access to a room.
-    pub fn allow(&mut self, room_name: &str, fingerprint: &str) {
-        if let Some(ref mut acl) = self.acl {
-            acl.entry(room_name.to_string())
+    pub fn allow(&self, room_name: &str, fingerprint: &str) {
+        if let Some(ref acl) = self.acl {
+            acl.lock().unwrap()
+                .entry(room_name.to_string())
                 .or_default()
                 .insert(fingerprint.to_string());
         }
@@ -251,6 +401,7 @@ impl RoomManager {
             (None, _) => true, // no ACL = open
             (Some(_), None) => false, // ACL enabled but no fingerprint
             (Some(acl), Some(fp)) => {
+                let acl = acl.lock().unwrap();
                 // Room not in ACL = open room (allow anyone authenticated)
                 match acl.get(room_name) {
                     None => true,
@@ -262,7 +413,7 @@ impl RoomManager {
 
     /// Join a room. Returns (participant_id, room_update_msg, all_senders) for broadcasting.
     pub fn join(
-        &mut self,
+        &self,
         room_name: &str,
         addr: std::net::SocketAddr,
         sender: ParticipantSender,
@@ -273,24 +424,25 @@ impl RoomManager {
             warn!(room = room_name, fingerprint = ?fingerprint, "unauthorized room join attempt");
             return Err("not authorized for this room".to_string());
         }
-        let was_empty = !self.rooms.contains_key(room_name)
-            || self.rooms.get(room_name).map_or(true, |r| r.is_empty());
-        let room = self.rooms.entry(room_name.to_string()).or_insert_with(Room::new);
+        let was_empty = self.rooms.get(room_name).map_or(true, |r| r.is_empty());
+        let mut room = self.rooms.entry(room_name.to_string()).or_insert_with(Room::new);
         let id = room.add(addr, sender, fingerprint.map(|s| s.to_string()), alias.map(|s| s.to_string()));
-        if was_empty {
-            let _ = self.event_tx.send(RoomEvent::LocalJoin { room: room_name.to_string() });
-        }
+        room.qualities.insert(id, ParticipantQuality::new());
         let update = wzp_proto::SignalMessage::RoomUpdate {
             count: room.len() as u32,
             participants: room.participant_list(),
         };
         let senders = room.all_senders();
+        drop(room); // release DashMap guard before event_tx send (not async, but good practice)
+        if was_empty {
+            let _ = self.event_tx.send(RoomEvent::LocalJoin { room: room_name.to_string() });
+        }
         Ok((id, update, senders))
     }
 
     /// Join a room via WebSocket. Convenience wrapper around `join()`.
     pub fn join_ws(
-        &mut self,
+        &self,
         room_name: &str,
         addr: std::net::SocketAddr,
         sender: tokio::sync::mpsc::Sender<Bytes>,
@@ -302,7 +454,7 @@ impl RoomManager {
 
     /// Get list of active room names.
     pub fn active_rooms(&self) -> Vec<String> {
-        self.rooms.keys().cloned().collect()
+        self.rooms.iter().map(|r| r.key().clone()).collect()
     }
 
     /// Get participant list for a room (fingerprint + alias).
@@ -322,24 +474,29 @@ impl RoomManager {
     }
 
     /// Leave a room. Returns (room_update_msg, remaining_senders) for broadcasting, or None if room is now empty.
-    pub fn leave(&mut self, room_name: &str, participant_id: ParticipantId) -> Option<(wzp_proto::SignalMessage, Vec<ParticipantSender>)> {
-        if let Some(room) = self.rooms.get_mut(room_name) {
-            room.remove(participant_id);
-            if room.is_empty() {
-                self.rooms.remove(room_name);
-                let _ = self.event_tx.send(RoomEvent::LocalLeave { room: room_name.to_string() });
-                info!(room = room_name, "room closed (empty)");
-                return None;
+    pub fn leave(&self, room_name: &str, participant_id: ParticipantId) -> Option<(wzp_proto::SignalMessage, Vec<ParticipantSender>)> {
+        let result = {
+            if let Some(mut room) = self.rooms.get_mut(room_name) {
+                room.qualities.remove(&participant_id);
+                room.remove(participant_id);
+                if room.is_empty() {
+                    drop(room); // release write guard before remove
+                    self.rooms.remove(room_name);
+                    let _ = self.event_tx.send(RoomEvent::LocalLeave { room: room_name.to_string() });
+                    info!(room = room_name, "room closed (empty)");
+                    return None;
+                }
+                let update = wzp_proto::SignalMessage::RoomUpdate {
+                    count: room.len() as u32,
+                    participants: room.participant_list(),
+                };
+                let senders = room.all_senders();
+                Some((update, senders))
+            } else {
+                None
             }
-            let update = wzp_proto::SignalMessage::RoomUpdate {
-                count: room.len() as u32,
-                participants: room.participant_list(),
-            };
-            let senders = room.all_senders();
-            Some((update, senders))
-        } else {
-            None
-        }
+        };
+        result
     }
 
     /// Get senders for all OTHER participants in a room.
@@ -359,9 +516,62 @@ impl RoomManager {
         self.rooms.get(room_name).map(|r| r.len()).unwrap_or(0)
     }
 
+    /// Check if a room exists and has participants.
+    pub fn is_room_active(&self, room_name: &str) -> bool {
+        self.rooms.contains_key(room_name)
+    }
+
     /// List all rooms with their sizes.
     pub fn list(&self) -> Vec<(String, usize)> {
-        self.rooms.iter().map(|(k, v)| (k.clone(), v.len())).collect()
+        self.rooms.iter().map(|r| (r.key().clone(), r.len())).collect()
+    }
+
+    /// Feed a quality report from a participant. If the room-wide weakest
+    /// tier changes, returns `(QualityDirective signal, all senders)` for
+    /// broadcasting.
+    pub fn observe_quality(
+        &self,
+        room_name: &str,
+        participant_id: ParticipantId,
+        report: &wzp_proto::packet::QualityReport,
+    ) -> Option<(wzp_proto::SignalMessage, Vec<ParticipantSender>)> {
+        let mut room = self.rooms.get_mut(room_name)?;
+
+        let tier_changed = room.qualities
+            .get_mut(&participant_id)
+            .and_then(|pq| pq.observe(report))
+            .is_some();
+
+        if !tier_changed {
+            return None;
+        }
+
+        // Compute the weakest tier across all participants in this room
+        let weakest = weakest_tier(room.qualities.values());
+
+        if weakest == room.current_tier {
+            return None;
+        }
+
+        // Room-wide tier changed -- update and broadcast directive
+        let old_tier = room.current_tier;
+        room.current_tier = weakest;
+        let profile = weakest.profile();
+        info!(
+            room = room_name,
+            old_tier = ?old_tier,
+            new_tier = ?weakest,
+            codec = ?profile.codec,
+            fec_ratio = profile.fec_ratio,
+            "room quality directive"
+        );
+
+        let directive = wzp_proto::SignalMessage::QualityDirective {
+            recommended_profile: profile,
+            reason: Some(format!("weakest link: {weakest:?}")),
+        };
+        let senders = room.all_senders();
+        Some((directive, senders))
     }
 }
 
@@ -382,18 +592,32 @@ impl TrunkedForwarder {
     /// Create a new trunked forwarder.
     ///
     /// `session_id` tags every entry pushed into the batcher so the receiver
-    /// can demultiplex packets by session.
+    /// can demultiplex packets by session. The batcher's `max_bytes` is
+    /// initialized from the transport's current PMTUD-discovered MTU so that
+    /// trunk frames fill the largest datagram the path supports (instead of
+    /// the conservative 1200-byte default).
     pub fn new(transport: Arc<wzp_transport::QuinnTransport>, session_id: [u8; 2]) -> Self {
+        let mut batcher = TrunkBatcher::new();
+        if let Some(mtu) = transport.max_datagram_size() {
+            batcher.max_bytes = mtu;
+        }
         Self {
             transport,
-            batcher: TrunkBatcher::new(),
+            batcher,
             session_id,
         }
     }
 
     /// Push a media packet into the batcher.  If the batcher is full it will
     /// flush automatically and the resulting trunk frame is sent immediately.
+    ///
+    /// Also refreshes `max_bytes` from the transport's PMTUD-discovered MTU
+    /// so the batcher fills larger datagrams as the path MTU grows.
     pub async fn send(&mut self, pkt: &wzp_proto::MediaPacket) -> anyhow::Result<()> {
+        // Refresh batcher limit from PMTUD (cheap: reads an atomic in quinn).
+        if let Some(mtu) = self.transport.max_datagram_size() {
+            self.batcher.max_bytes = mtu;
+        }
         let payload: Bytes = pkt.to_bytes();
         if let Some(frame) = self.batcher.push(self.session_id, payload) {
             self.send_frame(&frame)?;
@@ -430,7 +654,7 @@ impl TrunkedForwarder {
 /// into [`TrunkedForwarder`]s and flushed every 5 ms or when the batcher is
 /// full, reducing QUIC datagram overhead.
 pub async fn run_participant(
-    room_mgr: Arc<Mutex<RoomManager>>,
+    room_mgr: Arc<RoomManager>,
     room_name: String,
     participant_id: ParticipantId,
     transport: Arc<wzp_transport::QuinnTransport>,
@@ -456,7 +680,7 @@ pub async fn run_participant(
 
 /// Plain (non-trunked) forwarding loop — original behaviour.
 async fn run_participant_plain(
-    room_mgr: Arc<Mutex<RoomManager>>,
+    room_mgr: Arc<RoomManager>,
     room_name: String,
     participant_id: ParticipantId,
     transport: Arc<wzp_transport::QuinnTransport>,
@@ -474,6 +698,12 @@ async fn run_participant_plain(
     let mut send_errors = 0u64;
     let mut last_log_instant = std::time::Instant::now();
 
+    let mut tap_stats = if debug_tap.as_ref().map_or(false, |t| t.matches(&room_name)) {
+        Some(TapStats::new())
+    } else {
+        None
+    };
+
     info!(
         room = %room_name,
         participant = participant_id,
@@ -483,7 +713,6 @@ async fn run_participant_plain(
     );
 
     loop {
-        let recv_start = std::time::Instant::now();
         let pkt = match transport.recv_media().await {
             Ok(Some(pkt)) => pkt,
             Ok(None) => {
@@ -522,11 +751,16 @@ async fn run_participant_plain(
             metrics.update_session_quality(session_id, report);
         }
 
-        // Get current list of other participants
+        // Get current list of other participants + check quality directive
         let lock_start = std::time::Instant::now();
-        let others = {
-            let mgr = room_mgr.lock().await;
-            mgr.others(&room_name, participant_id)
+        let (others, quality_directive) = {
+            let directive = if let Some(ref report) = pkt.quality_report {
+                room_mgr.observe_quality(&room_name, participant_id, report)
+            } else {
+                None
+            };
+            let o = room_mgr.others(&room_name, participant_id);
+            (o, directive)
         };
         let lock_ms = lock_start.elapsed().as_millis() as u64;
         if lock_ms > 10 {
@@ -538,12 +772,25 @@ async fn run_participant_plain(
             );
         }
 
-        // Debug tap: log packet metadata
+        // Broadcast quality directive to all participants if tier changed
+        if let Some((directive, all_senders)) = quality_directive {
+            if let Some(ref tap) = debug_tap {
+                if tap.matches(&room_name) {
+                    tap.log_signal(&room_name, &directive);
+                }
+            }
+            broadcast_signal(&all_senders, &directive).await;
+        }
+
+        // Debug tap: log packet metadata + record stats
         if let Some(ref tap) = debug_tap {
             if tap.matches(&room_name) {
                 tap.log_packet(&room_name, "in", &addr, &pkt, others.len());
             }
         }
+        if let Some(ref mut ts) = tap_stats {
+            ts.record_in(&pkt, others.len());
+        }
 
         // Forward to all others
         let fwd_start = std::time::Instant::now();
@@ -601,10 +848,7 @@ async fn run_participant_plain(
 
         // Periodic stats log every 5 seconds
         if last_log_instant.elapsed() >= Duration::from_secs(5) {
-            let room_size = {
-                let mgr = room_mgr.lock().await;
-                mgr.room_size(&room_name)
-            };
+            let room_size = room_mgr.room_size(&room_name);
             info!(
                 room = %room_name,
                 participant = participant_id,
@@ -616,6 +860,10 @@ async fn run_participant_plain(
                 send_errors,
                 "participant stats"
             );
+            if let (Some(tap), Some(ts)) = (&debug_tap, &mut tap_stats) {
+                tap.log_stats(&room_name, ts);
+                ts.reset_period();
+            }
             max_recv_gap_ms = 0;
             max_forward_ms = 0;
             last_log_instant = std::time::Instant::now();
@@ -623,16 +871,28 @@ async fn run_participant_plain(
     }
 
     // Clean up — leave room and broadcast update to remaining participants
-    let mut mgr = room_mgr.lock().await;
-    if let Some((update, senders)) = mgr.leave(&room_name, participant_id) {
-        drop(mgr); // release lock before async broadcast
+    if let Some((update, senders)) = room_mgr.leave(&room_name, participant_id) {
+        if let Some(ref tap) = debug_tap {
+            if tap.matches(&room_name) {
+                tap.log_event(&room_name, "leave", &format!(
+                    "participant={participant_id} addr={addr} forwarded={packets_forwarded}"
+                ));
+                tap.log_signal(&room_name, &update);
+            }
+        }
         broadcast_signal(&senders, &update).await;
+    } else if let Some(ref tap) = debug_tap {
+        if tap.matches(&room_name) {
+            tap.log_event(&room_name, "leave", &format!(
+                "participant={participant_id} addr={addr} (room closed)"
+            ));
+        }
     }
 }
 
 /// Trunked forwarding loop — batches outgoing packets per peer.
 async fn run_participant_trunked(
-    room_mgr: Arc<Mutex<RoomManager>>,
+    room_mgr: Arc<RoomManager>,
     room_name: String,
     participant_id: ParticipantId,
     transport: Arc<wzp_transport::QuinnTransport>,
@@ -706,9 +966,14 @@ async fn run_participant_trunked(
                 }
 
                 let lock_start = std::time::Instant::now();
-                let others = {
-                    let mgr = room_mgr.lock().await;
-                    mgr.others(&room_name, participant_id)
+                let (others, quality_directive) = {
+                    let directive = if let Some(ref report) = pkt.quality_report {
+                        room_mgr.observe_quality(&room_name, participant_id, report)
+                    } else {
+                        None
+                    };
+                    let o = room_mgr.others(&room_name, participant_id);
+                    (o, directive)
                 };
                 let lock_ms = lock_start.elapsed().as_millis() as u64;
                 if lock_ms > 10 {
@@ -720,6 +985,11 @@ async fn run_participant_trunked(
                     );
                 }
 
+                // Broadcast quality directive to all participants if tier changed
+                if let Some((directive, all_senders)) = quality_directive {
+                    broadcast_signal(&all_senders, &directive).await;
+                }
+
                 let fwd_start = std::time::Instant::now();
                 let pkt_bytes = pkt.payload.len() as u64;
                 for other in &others {
@@ -768,10 +1038,7 @@ async fn run_participant_trunked(
 
                 // Periodic stats every 5 seconds
                 if last_log_instant.elapsed() >= Duration::from_secs(5) {
-                    let room_size = {
-                        let mgr = room_mgr.lock().await;
-                        mgr.room_size(&room_name)
-                    };
+                    let room_size = room_mgr.room_size(&room_name);
                     info!(
                         room = %room_name,
                         participant = participant_id,
@@ -812,9 +1079,7 @@ async fn run_participant_trunked(
         let _ = fwd.flush().await;
     }
 
-    let mut mgr = room_mgr.lock().await;
-    if let Some((update, senders)) = mgr.leave(&room_name, participant_id) {
-        drop(mgr);
+    if let Some((update, senders)) = room_mgr.leave(&room_name, participant_id) {
         broadcast_signal(&senders, &update).await;
     }
 }
@@ -838,7 +1103,7 @@ mod tests {
 
     #[test]
     fn room_join_leave() {
-        let mut mgr = RoomManager::new();
+        let mgr = RoomManager::new();
         assert_eq!(mgr.room_size("test"), 0);
         assert!(mgr.list().is_empty());
     }
@@ -860,7 +1125,7 @@ mod tests {
 
     #[test]
     fn acl_restricts_to_allowed() {
-        let mut mgr = RoomManager::with_acl();
+        let mgr = RoomManager::with_acl();
         mgr.allow("room1", "alice");
         mgr.allow("room1", "bob");
         assert!(mgr.is_authorized("room1", Some("alice")));
@@ -960,4 +1225,47 @@ mod tests {
         // Batcher should now be empty — nothing to flush.
         assert!(batcher.flush().is_none());
     }
+
+    fn make_report(loss_pct_f: f32, rtt_ms: u16) -> wzp_proto::packet::QualityReport {
+        wzp_proto::packet::QualityReport {
+            loss_pct: (loss_pct_f / 100.0 * 255.0) as u8,
+            rtt_4ms: (rtt_ms / 4) as u8,
+            jitter_ms: 10,
+            bitrate_cap_kbps: 200,
+        }
+    }
+
+    #[test]
+    fn participant_quality_starts_good() {
+        let pq = ParticipantQuality::new();
+        assert_eq!(pq.current_tier, Tier::Good);
+    }
+
+    #[test]
+    fn participant_quality_degrades_on_bad_reports() {
+        let mut pq = ParticipantQuality::new();
+        let bad = make_report(50.0, 300);
+        // Feed enough bad reports to trigger downgrade (3 consecutive)
+        for _ in 0..5 {
+            pq.observe(&bad);
+        }
+        assert_ne!(pq.current_tier, Tier::Good, "should degrade from Good");
+    }
+
+    #[test]
+    fn weakest_tier_picks_worst() {
+        let good = ParticipantQuality::new();
+        // good stays at Good tier
+
+        let mut bad = ParticipantQuality::new();
+        let bad_report = make_report(50.0, 300);
+        for _ in 0..5 {
+            bad.observe(&bad_report);
+        }
+        // bad should be degraded or catastrophic
+
+        let participants = vec![good, bad];
+        let weakest = weakest_tier(participants.iter());
+        assert_ne!(weakest, Tier::Good, "weakest should not be Good when one participant is bad");
+    }
 }

crates/wzp-relay/src/signal_hub.rs

@@ -7,7 +7,7 @@ use std::collections::HashMap;
 use std::sync::Arc;
 use std::time::Instant;
 
-use tracing::{info, warn};
+use tracing::info;
 use wzp_proto::{MediaTransport, SignalMessage};
 use wzp_transport::QuinnTransport;
 
@@ -86,6 +86,26 @@ impl SignalHub {
     pub fn alias(&self, fp: &str) -> Option<&str> {
         self.clients.get(fp).and_then(|c| c.alias.as_deref())
     }
+
+    /// Build a PresenceList message with all online users.
+    pub fn presence_list(&self) -> SignalMessage {
+        let users: Vec<wzp_proto::PresenceUser> = self
+            .clients
+            .values()
+            .map(|c| wzp_proto::PresenceUser {
+                fingerprint: c.fingerprint.clone(),
+                alias: c.alias.clone(),
+            })
+            .collect();
+        SignalMessage::PresenceList { users }
+    }
+
+    /// Broadcast a message to ALL connected signal clients.
+    pub async fn broadcast(&self, msg: &SignalMessage) {
+        for client in self.clients.values() {
+            let _ = client.transport.send_signal(msg).await;
+        }
+    }
 }
 
 #[cfg(test)]
@@ -94,7 +114,7 @@ mod tests {
 
     #[test]
     fn register_unregister() {
-        let mut hub = SignalHub::new();
+        let hub = SignalHub::new();
         assert_eq!(hub.online_count(), 0);
         assert!(!hub.is_online("alice"));
 

crates/wzp-relay/src/ws.rs

@@ -31,7 +31,7 @@ use crate::session_mgr::SessionManager;
 /// Shared state for WebSocket handlers.
 #[derive(Clone)]
 pub struct WsState {
-    pub room_mgr: Arc<Mutex<RoomManager>>,
+    pub room_mgr: Arc<RoomManager>,
     pub session_mgr: Arc<Mutex<SessionManager>>,
     pub auth_url: Option<String>,
     pub metrics: Arc<RelayMetrics>,
@@ -143,10 +143,9 @@ async fn handle_ws_connection(socket: WebSocket, room: String, state: WsState) {
     // 4. Join room with WS sender
     let addr: SocketAddr = ([0, 0, 0, 0], 0).into();
     let participant_id = {
-        let mut mgr = state.room_mgr.lock().await;
-        match mgr.join_ws(&room, addr, tx, fingerprint.as_deref()) {
+        match state.room_mgr.join_ws(&room, addr, tx, fingerprint.as_deref()) {
             Ok(id) => {
-                state.metrics.active_rooms.set(mgr.list().len() as i64);
+                state.metrics.active_rooms.set(state.room_mgr.list().len() as i64);
                 id
             }
             Err(e) => {
@@ -184,10 +183,7 @@ async fn handle_ws_connection(socket: WebSocket, room: String, state: WsState) {
     loop {
         match ws_rx.next().await {
             Some(Ok(Message::Binary(data))) => {
-                let others = {
-                    let mgr = state.room_mgr.lock().await;
-                    mgr.others(&room, participant_id)
-                };
+                let others = state.room_mgr.others(&room, participant_id);
                 for other in &others {
                     let _ = other.send_raw(&data).await;
                 }
@@ -214,11 +210,8 @@ async fn handle_ws_connection(socket: WebSocket, room: String, state: WsState) {
         reg.unregister_local(fp);
     }
 
-    {
-        let mut mgr = state.room_mgr.lock().await;
-        mgr.leave(&room, participant_id);
-        state.metrics.active_rooms.set(mgr.list().len() as i64);
-    }
+    state.room_mgr.leave(&room, participant_id);
+    state.metrics.active_rooms.set(state.room_mgr.list().len() as i64);
 
     let session_id_str: String = session_id.iter().map(|b| format!("{b:02x}")).collect();
     state.metrics.remove_session_metrics(&session_id_str);

crates/wzp-relay/tests/cross_relay_direct_call.rs

@@ -0,0 +1,321 @@
+//! Phase 4 integration test for cross-relay direct calling
+//! (PRD: .taskmaster/docs/prd_phase4_cross_relay_p2p.txt).
+//!
+//! Drives the call-registry cross-wiring + a simulated federation
+//! forward without spinning up actual relay binaries. The real
+//! main-loop and dispatcher code are exercised end-to-end in
+//! `reflect.rs` / `hole_punching.rs` already; this file focuses on
+//! the *new* invariants Phase 4 adds:
+//!
+//! 1. When Relay A forwards a DirectCallOffer, its local registry
+//!    stashes caller_reflexive_addr and leaves peer_relay_fp
+//!    unset (broadcast, answer-side will identify itself).
+//! 2. When Relay B's cross-relay dispatcher receives the forward,
+//!    its local registry stores the call with
+//!    peer_relay_fp = Some(relay_a_tls_fp).
+//! 3. When Relay B processes the local callee's answer, it sees
+//!    peer_relay_fp.is_some() and MUST NOT deliver the answer via
+//!    local signal_hub — instead it routes through federation.
+//! 4. When Relay A receives the forwarded answer via its
+//!    cross-relay dispatcher, it stashes callee_reflexive_addr
+//!    and emits a CallSetup to its local caller with
+//!    peer_direct_addr = callee_addr.
+//! 5. Final state: Alice's CallSetup carries Bob's reflex addr,
+//!    Bob's CallSetup carries Alice's reflex addr — cross-wired
+//!    through two relays + a federation link.
+
+use wzp_proto::{CallAcceptMode, SignalMessage};
+use wzp_relay::call_registry::CallRegistry;
+
+// ────────────────────────────────────────────────────────────────
+// Simulated dispatch helpers — these reproduce the exact logic
+// in main.rs without the tokio + federation boilerplate.
+// ────────────────────────────────────────────────────────────────
+
+const RELAY_A_TLS_FP: &str = "relay-A-tls-fingerprint";
+const RELAY_B_TLS_FP: &str = "relay-B-tls-fingerprint";
+const ALICE_ADDR: &str = "192.0.2.1:4433";
+const BOB_ADDR: &str = "198.51.100.9:4433";
+const RELAY_A_ADDR: &str = "203.0.113.5:4433";
+const RELAY_B_ADDR: &str = "203.0.113.10:4433";
+
+/// Helper that Alice's place_call sends.
+fn alice_offer(call_id: &str) -> SignalMessage {
+    SignalMessage::DirectCallOffer {
+        caller_fingerprint: "alice".into(),
+        caller_alias: None,
+        target_fingerprint: "bob".into(),
+        call_id: call_id.into(),
+        identity_pub: [0; 32],
+        ephemeral_pub: [0; 32],
+        signature: vec![],
+        supported_profiles: vec![],
+        caller_reflexive_addr: Some(ALICE_ADDR.into()),
+        caller_local_addrs: Vec::new(),
+        caller_mapped_addr: None,
+        caller_build_version: None,
+    }
+}
+
+/// Relay A receives Alice's offer. Target Bob is not local.
+/// Relay A wraps + broadcasts over federation, stashes the call
+/// locally with peer_relay_fp = None (broadcast — answer-side
+/// identifies itself).
+fn relay_a_handle_offer(reg_a: &mut CallRegistry, offer: &SignalMessage) -> SignalMessage {
+    match offer {
+        SignalMessage::DirectCallOffer {
+            caller_fingerprint,
+            target_fingerprint,
+            call_id,
+            caller_reflexive_addr,
+            ..
+        } => {
+            reg_a.create_call(
+                call_id.clone(),
+                caller_fingerprint.clone(),
+                target_fingerprint.clone(),
+            );
+            reg_a.set_caller_reflexive_addr(call_id, caller_reflexive_addr.clone());
+            // peer_relay_fp stays None — we don't know which peer
+            // will respond yet.
+        }
+        _ => panic!("not an offer"),
+    }
+    // Build the federation envelope the main loop would
+    // broadcast.
+    SignalMessage::FederatedSignalForward {
+        inner: Box::new(offer.clone()),
+        origin_relay_fp: RELAY_A_TLS_FP.into(),
+    }
+}
+
+/// Relay B receives a FederatedSignalForward(DirectCallOffer).
+/// This is the cross-relay dispatcher task code in main.rs —
+/// reproduced here for the test.
+fn relay_b_handle_forwarded_offer(reg_b: &mut CallRegistry, forward: &SignalMessage) {
+    let (inner, origin_relay_fp) = match forward {
+        SignalMessage::FederatedSignalForward { inner, origin_relay_fp } => {
+            (inner.as_ref().clone(), origin_relay_fp.clone())
+        }
+        _ => panic!("not a forward"),
+    };
+    // Loop-prevention: drop self-sourced.
+    assert_ne!(origin_relay_fp, RELAY_B_TLS_FP);
+
+    let SignalMessage::DirectCallOffer {
+        caller_fingerprint,
+        target_fingerprint,
+        call_id,
+        caller_reflexive_addr,
+        ..
+    } = inner
+    else {
+        panic!("inner was not DirectCallOffer");
+    };
+
+    // Simulated: target is local to B (Bob is registered here).
+    reg_b.create_call(
+        call_id.clone(),
+        caller_fingerprint,
+        target_fingerprint,
+    );
+    reg_b.set_caller_reflexive_addr(&call_id, caller_reflexive_addr);
+    reg_b.set_peer_relay_fp(&call_id, Some(origin_relay_fp));
+}
+
+/// Bob's answer — AcceptTrusted with his reflex addr.
+fn bob_answer(call_id: &str) -> SignalMessage {
+    SignalMessage::DirectCallAnswer {
+        call_id: call_id.into(),
+        accept_mode: CallAcceptMode::AcceptTrusted,
+        identity_pub: None,
+        ephemeral_pub: None,
+        signature: None,
+        chosen_profile: None,
+        callee_reflexive_addr: Some(BOB_ADDR.into()),
+        callee_local_addrs: Vec::new(),
+        callee_mapped_addr: None,
+        callee_build_version: None,
+    }
+}
+
+/// Relay B handles the LOCAL callee's answer. If peer_relay_fp
+/// is Some, wrap the answer in a FederatedSignalForward + emit the
+/// local CallSetup to Bob. Returns the (forward_envelope,
+/// bob_call_setup) pair.
+fn relay_b_handle_local_answer(
+    reg_b: &mut CallRegistry,
+    answer: &SignalMessage,
+) -> (SignalMessage, SignalMessage) {
+    let (call_id, mode, callee_addr) = match answer {
+        SignalMessage::DirectCallAnswer {
+            call_id,
+            accept_mode,
+            callee_reflexive_addr,
+            ..
+        } => (call_id.clone(), *accept_mode, callee_reflexive_addr.clone()),
+        _ => panic!(),
+    };
+    // Stash callee addr + activate.
+    reg_b.set_active(&call_id, mode, format!("call-{call_id}"));
+    reg_b.set_callee_reflexive_addr(&call_id, callee_addr);
+    let call = reg_b.get(&call_id).unwrap();
+    let caller_addr = call.caller_reflexive_addr.clone();
+    let callee_addr = call.callee_reflexive_addr.clone();
+    assert!(
+        call.peer_relay_fp.is_some(),
+        "Relay B must know this call is cross-relay"
+    );
+
+    // Forward the answer back over federation.
+    let forward = SignalMessage::FederatedSignalForward {
+        inner: Box::new(answer.clone()),
+        origin_relay_fp: RELAY_B_TLS_FP.into(),
+    };
+
+    // Local CallSetup for Bob — peer_direct_addr = Alice's addr.
+    let setup_for_bob = SignalMessage::CallSetup {
+        call_id: call_id.clone(),
+        room: format!("call-{call_id}"),
+        relay_addr: RELAY_B_ADDR.into(),
+        peer_direct_addr: caller_addr,
+        peer_local_addrs: Vec::new(),
+        peer_mapped_addr: None,
+    };
+    let _ = callee_addr;
+    (forward, setup_for_bob)
+}
+
+/// Relay A's cross-relay dispatcher receives the forwarded answer.
+/// It stashes the callee addr, forwards the raw answer to local
+/// Alice, and emits a CallSetup with peer_direct_addr = Bob's addr.
+fn relay_a_handle_forwarded_answer(
+    reg_a: &mut CallRegistry,
+    forward: &SignalMessage,
+) -> SignalMessage {
+    let (inner, origin_relay_fp) = match forward {
+        SignalMessage::FederatedSignalForward { inner, origin_relay_fp } => {
+            (inner.as_ref().clone(), origin_relay_fp.clone())
+        }
+        _ => panic!("not a forward"),
+    };
+    assert_ne!(origin_relay_fp, RELAY_A_TLS_FP);
+
+    let SignalMessage::DirectCallAnswer {
+        call_id,
+        accept_mode,
+        callee_reflexive_addr,
+        ..
+    } = inner
+    else {
+        panic!("inner was not DirectCallAnswer");
+    };
+    assert_eq!(accept_mode, CallAcceptMode::AcceptTrusted);
+
+    reg_a.set_active(&call_id, accept_mode, format!("call-{call_id}"));
+    reg_a.set_callee_reflexive_addr(&call_id, callee_reflexive_addr.clone());
+
+    // Alice's CallSetup — peer_direct_addr = Bob's addr.
+    SignalMessage::CallSetup {
+        call_id: call_id.clone(),
+        room: format!("call-{call_id}"),
+        relay_addr: RELAY_A_ADDR.into(),
+        peer_direct_addr: callee_reflexive_addr,
+        peer_local_addrs: Vec::new(),
+        peer_mapped_addr: None,
+    }
+}
+
+// ────────────────────────────────────────────────────────────────
+// Tests
+// ────────────────────────────────────────────────────────────────
+
+#[test]
+fn cross_relay_offer_forwards_and_stashes_peer_relay_fp() {
+    let mut reg_a = CallRegistry::new();
+    let mut reg_b = CallRegistry::new();
+
+    let offer = alice_offer("c-xrelay-1");
+    let forward = relay_a_handle_offer(&mut reg_a, &offer);
+
+    // Relay A's local view: call exists, caller addr stashed,
+    // peer_relay_fp still None (broadcast — answer identifies the
+    // peer).
+    let call_a = reg_a.get("c-xrelay-1").unwrap();
+    assert_eq!(call_a.caller_fingerprint, "alice");
+    assert_eq!(call_a.callee_fingerprint, "bob");
+    assert_eq!(call_a.caller_reflexive_addr.as_deref(), Some(ALICE_ADDR));
+    assert!(call_a.peer_relay_fp.is_none());
+
+    // Relay B dispatches the forward: creates the call locally
+    // and stashes peer_relay_fp = Relay A.
+    relay_b_handle_forwarded_offer(&mut reg_b, &forward);
+    let call_b = reg_b.get("c-xrelay-1").unwrap();
+    assert_eq!(call_b.caller_fingerprint, "alice");
+    assert_eq!(call_b.callee_fingerprint, "bob");
+    assert_eq!(call_b.caller_reflexive_addr.as_deref(), Some(ALICE_ADDR));
+    assert_eq!(call_b.peer_relay_fp.as_deref(), Some(RELAY_A_TLS_FP));
+}
+
+#[test]
+fn cross_relay_answer_crosswires_peer_direct_addrs() {
+    let mut reg_a = CallRegistry::new();
+    let mut reg_b = CallRegistry::new();
+
+    // Full round trip: offer → forward → dispatch → answer →
+    // forward back → dispatch → both CallSetups.
+    let offer = alice_offer("c-xrelay-2");
+    let offer_forward = relay_a_handle_offer(&mut reg_a, &offer);
+    relay_b_handle_forwarded_offer(&mut reg_b, &offer_forward);
+
+    // Bob answers on Relay B.
+    let answer = bob_answer("c-xrelay-2");
+    let (answer_forward, setup_for_bob) =
+        relay_b_handle_local_answer(&mut reg_b, &answer);
+
+    // Bob's CallSetup carries Alice's addr.
+    match setup_for_bob {
+        SignalMessage::CallSetup { peer_direct_addr, relay_addr, .. } => {
+            assert_eq!(peer_direct_addr.as_deref(), Some(ALICE_ADDR));
+            assert_eq!(relay_addr, RELAY_B_ADDR);
+        }
+        _ => panic!("wrong variant"),
+    }
+
+    // Alice's dispatcher receives the forwarded answer and builds
+    // her CallSetup.
+    let setup_for_alice = relay_a_handle_forwarded_answer(&mut reg_a, &answer_forward);
+    match setup_for_alice {
+        SignalMessage::CallSetup { peer_direct_addr, relay_addr, .. } => {
+            assert_eq!(peer_direct_addr.as_deref(), Some(BOB_ADDR));
+            assert_eq!(relay_addr, RELAY_A_ADDR);
+        }
+        _ => panic!("wrong variant"),
+    }
+
+    // Both registries agree on caller + callee reflex addrs after
+    // the full round-trip.
+    for reg in [&reg_a, &reg_b] {
+        let c = reg.get("c-xrelay-2").unwrap();
+        assert_eq!(c.caller_reflexive_addr.as_deref(), Some(ALICE_ADDR));
+        assert_eq!(c.callee_reflexive_addr.as_deref(), Some(BOB_ADDR));
+    }
+}
+
+#[test]
+fn cross_relay_loop_prevention_drops_self_sourced_forward() {
+    // A FederatedSignalForward that circles back to the origin
+    // relay should be dropped before it hits the call registry.
+    let forward = SignalMessage::FederatedSignalForward {
+        inner: Box::new(alice_offer("c-loop")),
+        origin_relay_fp: RELAY_B_TLS_FP.into(),
+    };
+    // The dispatcher in main.rs calls this explicit check before
+    // doing any work. Reproduce it inline.
+    let origin = match &forward {
+        SignalMessage::FederatedSignalForward { origin_relay_fp, .. } => origin_relay_fp.clone(),
+        _ => unreachable!(),
+    };
+    // Relay B sees origin == its own fp → drop.
+    assert_eq!(origin, RELAY_B_TLS_FP, "loop-prevention triggers on self-fp");
+}

crates/wzp-relay/tests/federation.rs

@@ -0,0 +1,662 @@
+//! Tests for `wzp_relay::federation`.
+//!
+//! Covers:
+//!  - room_hash determinism and uniqueness
+//!  - is_global_room (static config + call-* implicit global)
+//!  - resolve_global_room
+//!  - global_room_hash
+//!  - forward_to_peers with zero peers (no-op)
+//!  - forward_to_peers with live QUIC peer links
+//!  - broadcast_signal to live QUIC peers
+//!  - send_signal_to_peer targeted routing
+//!  - find_peer_by_fingerprint / find_peer_by_addr / check_inbound_trust
+//!  - set_cross_relay_tx + local_tls_fp accessors
+
+use std::collections::HashSet;
+use std::net::{Ipv4Addr, SocketAddr};
+use std::sync::Arc;
+use std::time::Duration;
+
+use bytes::Bytes;
+use wzp_proto::{MediaTransport, SignalMessage};
+use wzp_relay::config::{PeerConfig, TrustedConfig};
+use wzp_relay::event_log::EventLogger;
+use wzp_relay::federation::{room_hash, FederationManager};
+use wzp_relay::metrics::RelayMetrics;
+use wzp_relay::room::RoomManager;
+use wzp_transport::{client_config, create_endpoint, server_config, QuinnTransport};
+
+// ───────────────────────────── helpers ──────────────────────────────
+
+/// Create a FederationManager for unit tests (no live peers).
+fn create_test_fm(global_rooms: HashSet<String>) -> Arc<FederationManager> {
+    create_test_fm_full(vec![], vec![], global_rooms)
+}
+
+/// Create a FederationManager with full config (peers + trusted + global rooms).
+fn create_test_fm_full(
+    peers: Vec<PeerConfig>,
+    trusted: Vec<TrustedConfig>,
+    global_rooms: HashSet<String>,
+) -> Arc<FederationManager> {
+    let _ = rustls::crypto::ring::default_provider().install_default();
+    let (sc, _cert) = server_config();
+    let ep = create_endpoint((Ipv4Addr::LOCALHOST, 0).into(), Some(sc))
+        .expect("test endpoint");
+    let room_mgr = Arc::new(RoomManager::new());
+    let metrics = Arc::new(RelayMetrics::new());
+    let event_log = EventLogger::Noop;
+
+    Arc::new(FederationManager::new(
+        peers,
+        trusted,
+        global_rooms,
+        room_mgr,
+        ep,
+        "test-relay-fp-abc123".into(),
+        metrics,
+        event_log,
+    ))
+}
+
+/// Build an in-process QUIC client/server pair on loopback.
+/// Returns (client_transport, server_transport, endpoints).
+/// The endpoints must be kept alive for the test duration.
+async fn connected_pair() -> (
+    Arc<QuinnTransport>,
+    Arc<QuinnTransport>,
+    (quinn::Endpoint, quinn::Endpoint),
+) {
+    let _ = rustls::crypto::ring::default_provider().install_default();
+
+    let (sc, _cert_der) = server_config();
+    let server_addr: SocketAddr = (Ipv4Addr::LOCALHOST, 0).into();
+    let server_ep = create_endpoint(server_addr, Some(sc)).expect("server endpoint");
+    let server_listen = server_ep.local_addr().expect("server local addr");
+
+    let client_bind: SocketAddr = (Ipv4Addr::LOCALHOST, 0).into();
+    let client_ep = create_endpoint(client_bind, None).expect("client endpoint");
+
+    let server_ep_clone = server_ep.clone();
+    let accept_fut = tokio::spawn(async move {
+        let conn = wzp_transport::accept(&server_ep_clone)
+            .await
+            .expect("accept");
+        Arc::new(QuinnTransport::new(conn))
+    });
+
+    let client_conn =
+        wzp_transport::connect(&client_ep, server_listen, "localhost", client_config())
+            .await
+            .expect("connect");
+    let client_transport = Arc::new(QuinnTransport::new(client_conn));
+    let server_transport = accept_fut.await.expect("join accept task");
+
+    (client_transport, server_transport, (server_ep, client_ep))
+}
+
+// ───────────────────── 1. room_hash determinism ─────────────────────
+
+#[test]
+fn room_hash_deterministic() {
+    let h1 = room_hash("podcast");
+    let h2 = room_hash("podcast");
+    assert_eq!(h1, h2);
+}
+
+#[test]
+fn room_hash_different_rooms() {
+    let h1 = room_hash("room-a");
+    let h2 = room_hash("room-b");
+    assert_ne!(h1, h2);
+}
+
+#[test]
+fn room_hash_is_8_bytes() {
+    let h = room_hash("some-room");
+    assert_eq!(h.len(), 8);
+}
+
+#[test]
+fn room_hash_empty_string() {
+    // Should not panic on empty input
+    let h = room_hash("");
+    assert_eq!(h.len(), 8);
+    // And should differ from a non-empty room
+    assert_ne!(h, room_hash("nonempty"));
+}
+
+#[test]
+fn room_hash_case_sensitive() {
+    // "Podcast" and "podcast" are different rooms
+    let h1 = room_hash("Podcast");
+    let h2 = room_hash("podcast");
+    assert_ne!(h1, h2);
+}
+
+// ───────────────── 2. is_global_room / resolve_global_room ──────────
+
+#[tokio::test]
+async fn is_global_room_static_config() {
+    let global: HashSet<String> = ["podcast", "lobby"].iter().map(|s| s.to_string()).collect();
+    let fm = create_test_fm(global);
+
+    assert!(fm.is_global_room("podcast"));
+    assert!(fm.is_global_room("lobby"));
+    assert!(!fm.is_global_room("private-room"));
+    assert!(!fm.is_global_room(""));
+}
+
+#[tokio::test]
+async fn is_global_room_call_prefix_implicit() {
+    // Phase 4.1: call-* rooms are implicitly global
+    let fm = create_test_fm(HashSet::new());
+
+    assert!(fm.is_global_room("call-abc123"));
+    assert!(fm.is_global_room("call-"));
+    assert!(fm.is_global_room("call-some-uuid-here"));
+    // But not just "call" without the dash
+    assert!(!fm.is_global_room("call"));
+    assert!(!fm.is_global_room("callback"));
+}
+
+#[tokio::test]
+async fn resolve_global_room_static() {
+    let global: HashSet<String> = ["podcast"].iter().map(|s| s.to_string()).collect();
+    let fm = create_test_fm(global);
+
+    assert_eq!(fm.resolve_global_room("podcast"), Some("podcast".into()));
+    assert_eq!(fm.resolve_global_room("unknown"), None);
+}
+
+#[tokio::test]
+async fn resolve_global_room_call_prefix() {
+    let fm = create_test_fm(HashSet::new());
+
+    let resolved = fm.resolve_global_room("call-test-123");
+    assert_eq!(resolved, Some("call-test-123".into()));
+}
+
+#[tokio::test]
+async fn global_room_hash_uses_canonical_name() {
+    let global: HashSet<String> = ["podcast"].iter().map(|s| s.to_string()).collect();
+    let fm = create_test_fm(global);
+
+    // For a known global room, global_room_hash should match room_hash of the canonical name
+    let expected = room_hash("podcast");
+    assert_eq!(fm.global_room_hash("podcast"), expected);
+}
+
+#[tokio::test]
+async fn global_room_hash_unknown_room_falls_through() {
+    let fm = create_test_fm(HashSet::new());
+
+    // Unknown room: just hashes whatever was passed
+    let expected = room_hash("random-room");
+    assert_eq!(fm.global_room_hash("random-room"), expected);
+}
+
+#[tokio::test]
+async fn global_room_hash_call_prefix() {
+    let fm = create_test_fm(HashSet::new());
+
+    // call-* resolves to itself
+    let expected = room_hash("call-xyz");
+    assert_eq!(fm.global_room_hash("call-xyz"), expected);
+}
+
+// ───────────────── 3. forward_to_peers with zero peers ──────────────
+
+#[tokio::test]
+async fn forward_to_peers_empty_returns_immediately() {
+    let fm = create_test_fm(HashSet::new());
+    let hash = room_hash("room");
+    let data = Bytes::from_static(b"test-media-payload");
+
+    // Should not panic or hang
+    let result = tokio::time::timeout(
+        Duration::from_secs(2),
+        fm.forward_to_peers("room", &hash, &data),
+    )
+    .await;
+    assert!(result.is_ok(), "forward_to_peers should return immediately with no peers");
+}
+
+// ─────────── 4. forward_to_peers with live QUIC peer links ──────────
+
+#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
+async fn forward_to_peers_delivers_tagged_datagram() {
+    // We create a FederationManager and manually wire a connected QUIC
+    // pair to simulate a peer link. The fm holds the server-side
+    // transport; we read from the client side to verify delivery.
+    let fm = create_test_fm(HashSet::new());
+
+    let (client_transport, server_transport, _endpoints) = connected_pair().await;
+
+    // Manually insert a PeerLink by using handle_inbound's internal
+    // pattern: we call the private peer_links mutex directly. Since
+    // PeerLink is private, we instead use handle_inbound which calls
+    // run_federation_link. But that requires a full signal loop.
+    //
+    // Alternative approach: spawn a mock "federation relay" server,
+    // have the FM connect to it via connect_to_peer, and read back
+    // from the server side. But connect_to_peer also starts the full
+    // link loop.
+    //
+    // Simplest: create a second FM that acts as the peer, and use
+    // the broadcast_signal / forward_to_peers pattern after the link
+    // is established via handle_inbound.
+    //
+    // Actually the simplest approach for testing forward_to_peers is
+    // to accept that PeerLink is private, so we instead test through
+    // the full federation link lifecycle. We'll spawn a mini relay
+    // that does the FederationHello handshake and then reads datagrams.
+
+    // Approach: spawn the server side to do the hello exchange, then
+    // the fm handle_inbound will register the link, then we can call
+    // forward_to_peers and read from the server side... But
+    // handle_inbound blocks in run_federation_link.
+    //
+    // Final approach: we test the wire format directly. The client
+    // side is "us" (the relay) — we send a tagged datagram manually,
+    // and verify the peer side receives it with the correct format.
+    // This tests the same logic as forward_to_peers without needing
+    // peer_links access.
+
+    let room = "test-room";
+    let rh = room_hash(room);
+    let media = b"opus-frame-data-here";
+
+    // Build the tagged datagram the same way forward_to_peers does
+    let mut tagged = Vec::with_capacity(8 + media.len());
+    tagged.extend_from_slice(&rh);
+    tagged.extend_from_slice(media);
+
+    // Send from the server side (as if we are the relay forwarding)
+    server_transport
+        .send_raw_datagram(&tagged)
+        .expect("send datagram");
+
+    // Read from client side (as if we are the peer relay receiving)
+    let received = tokio::time::timeout(
+        Duration::from_secs(2),
+        client_transport.connection().read_datagram(),
+    )
+    .await
+    .expect("should receive within timeout")
+    .expect("read_datagram ok");
+
+    // Verify: first 8 bytes are the room hash, remainder is media
+    assert!(received.len() >= 8, "datagram too short");
+    let mut recv_hash = [0u8; 8];
+    recv_hash.copy_from_slice(&received[..8]);
+    assert_eq!(recv_hash, rh, "room hash mismatch");
+    assert_eq!(&received[8..], media, "media payload mismatch");
+
+    drop(client_transport);
+    drop(server_transport);
+}
+
+// ─────────── 5. broadcast_signal to live QUIC peers ─────────────────
+
+#[tokio::test(flavor = "multi_thread", worker_threads = 4)]
+async fn broadcast_signal_sends_to_all_peers() {
+    // We need the peer links to be registered inside the FM.
+    // The simplest approach: spawn a mock peer relay that accepts
+    // federation connections, does the FederationHello handshake,
+    // and then reads signals.
+
+    let _ = rustls::crypto::ring::default_provider().install_default();
+
+    // Create a mock "peer relay" server endpoint
+    let (sc, _cert) = server_config();
+    let peer_addr: SocketAddr = (Ipv4Addr::LOCALHOST, 0).into();
+    let peer_ep = create_endpoint(peer_addr, Some(sc)).expect("peer endpoint");
+    let peer_listen = peer_ep.local_addr().expect("peer local addr");
+
+    // The FM that will connect outbound
+    let peer_cfg = PeerConfig {
+        url: peer_listen.to_string(),
+        fingerprint: "aa:bb:cc:dd".into(),
+        label: Some("mock-peer".into()),
+    };
+    let global: HashSet<String> = ["podcast"].iter().map(|s| s.to_string()).collect();
+    let fm = create_test_fm_full(vec![peer_cfg], vec![], global);
+
+    // Spawn the FM's run (which will try to connect to our mock peer)
+    let fm_clone = fm.clone();
+    let _fm_task = tokio::spawn(async move {
+        fm_clone.run().await;
+    });
+
+    // Accept the connection on the mock peer side
+    let peer_ep_clone = peer_ep.clone();
+    let peer_transport = tokio::time::timeout(Duration::from_secs(5), async {
+        let conn = wzp_transport::accept(&peer_ep_clone).await.expect("accept");
+        Arc::new(QuinnTransport::new(conn))
+    })
+    .await
+    .expect("FM should connect to mock peer within 5s");
+
+    // The FM sends FederationHello as the first signal. Read it.
+    let hello = tokio::time::timeout(
+        Duration::from_secs(2),
+        peer_transport.recv_signal(),
+    )
+    .await
+    .expect("hello timeout")
+    .expect("recv ok")
+    .expect("some message");
+
+    match hello {
+        SignalMessage::FederationHello { tls_fingerprint } => {
+            assert_eq!(tls_fingerprint, "test-relay-fp-abc123");
+        }
+        other => panic!("expected FederationHello, got: {:?}", std::mem::discriminant(&other)),
+    }
+
+    // Now the FM's run_federation_link registered the peer in peer_links
+    // and will announce active global rooms. We may receive
+    // GlobalRoomActive signals next (for any rooms the FM has active).
+    // For this test, no local participants, so no GlobalRoomActive.
+
+    // Give the link time to fully set up
+    tokio::time::sleep(Duration::from_millis(100)).await;
+
+    // Now call broadcast_signal on the FM
+    let test_msg = SignalMessage::FederatedSignalForward {
+        inner: Box::new(SignalMessage::Reflect),
+        origin_relay_fp: "other-relay-fp".into(),
+    };
+    let count = fm.broadcast_signal(&test_msg).await;
+    assert_eq!(count, 1, "should have broadcast to exactly 1 peer");
+
+    // Read the signal on the peer side
+    let received = tokio::time::timeout(
+        Duration::from_secs(2),
+        peer_transport.recv_signal(),
+    )
+    .await
+    .expect("broadcast signal timeout")
+    .expect("recv ok")
+    .expect("some message");
+
+    match received {
+        SignalMessage::FederatedSignalForward { origin_relay_fp, .. } => {
+            assert_eq!(origin_relay_fp, "other-relay-fp");
+        }
+        other => panic!("expected FederatedSignalForward, got: {:?}", std::mem::discriminant(&other)),
+    }
+
+    drop(peer_transport);
+}
+
+// ──────────── 6. send_signal_to_peer targeted routing ───────────────
+
+#[tokio::test(flavor = "multi_thread", worker_threads = 4)]
+async fn send_signal_to_peer_unknown_fp_returns_error() {
+    let fm = create_test_fm(HashSet::new());
+
+    let msg = SignalMessage::Reflect;
+    let result = fm.send_signal_to_peer("nonexistent-fp", &msg).await;
+    assert!(result.is_err());
+    assert!(result.unwrap_err().contains("no active federation link"));
+}
+
+// ──────────── 7. find_peer_by_fingerprint / addr / trust ────────────
+
+#[tokio::test]
+async fn find_peer_by_fingerprint_matches() {
+    let peer = PeerConfig {
+        url: "10.0.0.1:4433".into(),
+        fingerprint: "AA:BB:CC:DD".into(),
+        label: Some("relay-eu".into()),
+    };
+    let fm = create_test_fm_full(vec![peer], vec![], HashSet::new());
+
+    // Normalized match (colons removed, lowercased)
+    let found = fm.find_peer_by_fingerprint("aabbccdd");
+    assert!(found.is_some());
+    assert_eq!(found.unwrap().label.as_deref(), Some("relay-eu"));
+
+    // With colons
+    let found2 = fm.find_peer_by_fingerprint("AA:BB:CC:DD");
+    assert!(found2.is_some());
+
+    // Non-matching
+    assert!(fm.find_peer_by_fingerprint("11:22:33:44").is_none());
+}
+
+#[tokio::test]
+async fn find_peer_by_addr_matches_ip() {
+    let peer = PeerConfig {
+        url: "10.0.0.1:4433".into(),
+        fingerprint: "aabb".into(),
+        label: None,
+    };
+    let fm = create_test_fm_full(vec![peer], vec![], HashSet::new());
+
+    // Same IP, different port still matches (find_peer_by_addr matches by IP)
+    let addr: SocketAddr = "10.0.0.1:9999".parse().unwrap();
+    let found = fm.find_peer_by_addr(addr);
+    assert!(found.is_some());
+
+    // Different IP
+    let addr2: SocketAddr = "10.0.0.2:4433".parse().unwrap();
+    assert!(fm.find_peer_by_addr(addr2).is_none());
+}
+
+#[tokio::test]
+async fn find_trusted_by_fingerprint() {
+    let trusted = TrustedConfig {
+        fingerprint: "AA:BB:CC:DD:EE".into(),
+        label: Some("trusted-relay".into()),
+    };
+    let fm = create_test_fm_full(vec![], vec![trusted], HashSet::new());
+
+    let found = fm.find_trusted_by_fingerprint("aabbccddee");
+    assert!(found.is_some());
+    assert_eq!(found.unwrap().label.as_deref(), Some("trusted-relay"));
+
+    assert!(fm.find_trusted_by_fingerprint("ffffffff").is_none());
+}
+
+#[tokio::test]
+async fn check_inbound_trust_prefers_peer_by_addr() {
+    let peer = PeerConfig {
+        url: "10.0.0.1:4433".into(),
+        fingerprint: "aabb".into(),
+        label: Some("peer-relay".into()),
+    };
+    let trusted = TrustedConfig {
+        fingerprint: "ccdd".into(),
+        label: Some("trusted-relay".into()),
+    };
+    let fm = create_test_fm_full(vec![peer], vec![trusted], HashSet::new());
+
+    // Matches by addr (peer takes priority)
+    let addr: SocketAddr = "10.0.0.1:5555".parse().unwrap();
+    let label = fm.check_inbound_trust(addr, "ccdd");
+    assert_eq!(label, Some("peer-relay".into()));
+}
+
+#[tokio::test]
+async fn check_inbound_trust_falls_back_to_trusted_fp() {
+    let trusted = TrustedConfig {
+        fingerprint: "CC:DD".into(),
+        label: Some("trusted-relay".into()),
+    };
+    let fm = create_test_fm_full(vec![], vec![trusted], HashSet::new());
+
+    // No peer matches, but trusted fingerprint matches
+    let addr: SocketAddr = "10.99.99.99:1234".parse().unwrap();
+    let label = fm.check_inbound_trust(addr, "ccdd");
+    assert_eq!(label, Some("trusted-relay".into()));
+}
+
+#[tokio::test]
+async fn check_inbound_trust_returns_none_for_unknown() {
+    let fm = create_test_fm(HashSet::new());
+    let addr: SocketAddr = "10.0.0.1:4433".parse().unwrap();
+    assert!(fm.check_inbound_trust(addr, "unknown-fp").is_none());
+}
+
+// ──────────── 8. set_cross_relay_tx + local_tls_fp ──────────────────
+
+#[tokio::test]
+async fn local_tls_fp_returns_configured_value() {
+    let fm = create_test_fm(HashSet::new());
+    assert_eq!(fm.local_tls_fp(), "test-relay-fp-abc123");
+}
+
+#[tokio::test]
+async fn set_cross_relay_tx_wires_channel() {
+    let fm = create_test_fm(HashSet::new());
+    let (tx, mut rx) = tokio::sync::mpsc::channel(16);
+
+    fm.set_cross_relay_tx(tx).await;
+
+    // The channel is now wired — we can't easily test it without
+    // going through handle_signal, but we can at least verify it
+    // doesn't panic and the fm accepted the sender.
+    // (The channel itself works — we test the Sender.)
+    let msg = SignalMessage::Reflect;
+    let _ = rx.try_recv(); // should be empty
+    drop(rx);
+}
+
+// ──────────── 9. broadcast_signal with zero peers ───────────────────
+
+#[tokio::test]
+async fn broadcast_signal_zero_peers_returns_zero() {
+    let fm = create_test_fm(HashSet::new());
+    let msg = SignalMessage::Reflect;
+    let count = fm.broadcast_signal(&msg).await;
+    assert_eq!(count, 0);
+}
+
+// ──────────── 10. get_remote_participants with no links ─────────────
+
+#[tokio::test]
+async fn get_remote_participants_empty_with_no_links() {
+    let fm = create_test_fm(HashSet::new());
+    let participants = fm.get_remote_participants("podcast").await;
+    assert!(participants.is_empty());
+}
+
+// ─────── 11. Federation media egress with live QUIC connection ──────
+
+#[tokio::test(flavor = "multi_thread", worker_threads = 4)]
+async fn federation_media_egress_forwards_to_peer() {
+    // This test verifies the full media path:
+    //   local media -> federation egress channel -> forward_to_peers -> peer reads datagram
+    //
+    // We set up a real QUIC federation link via fm.run() connecting to
+    // a mock peer, then push media through the room manager's federation
+    // egress channel.
+
+    let _ = rustls::crypto::ring::default_provider().install_default();
+
+    // Mock peer relay
+    let (sc, _cert) = server_config();
+    let peer_addr: SocketAddr = (Ipv4Addr::LOCALHOST, 0).into();
+    let peer_ep = create_endpoint(peer_addr, Some(sc)).expect("peer endpoint");
+    let peer_listen = peer_ep.local_addr().expect("peer local addr");
+
+    let peer_cfg = PeerConfig {
+        url: peer_listen.to_string(),
+        fingerprint: "ee:ff:00:11".into(),
+        label: Some("egress-peer".into()),
+    };
+    let global: HashSet<String> = ["podcast"].iter().map(|s| s.to_string()).collect();
+    let fm = create_test_fm_full(vec![peer_cfg], vec![], global);
+
+    // Start the FM (connects to mock peer)
+    let fm_clone = fm.clone();
+    let _fm_task = tokio::spawn(async move { fm_clone.run().await });
+
+    // Accept the connection
+    let peer_ep_clone = peer_ep.clone();
+    let peer_transport = tokio::time::timeout(Duration::from_secs(5), async {
+        let conn = wzp_transport::accept(&peer_ep_clone).await.expect("accept");
+        Arc::new(QuinnTransport::new(conn))
+    })
+    .await
+    .expect("FM should connect within 5s");
+
+    // Read the FederationHello
+    let _hello = tokio::time::timeout(
+        Duration::from_secs(2),
+        peer_transport.recv_signal(),
+    )
+    .await
+    .expect("hello timeout")
+    .expect("recv ok")
+    .expect("some message");
+
+    // Wait for link setup
+    tokio::time::sleep(Duration::from_millis(100)).await;
+
+    // Now send media via forward_to_peers
+    let room = "podcast";
+    let rh = room_hash(room);
+    let media_payload = Bytes::from_static(b"test-opus-frame-1234567890");
+
+    fm.forward_to_peers(room, &rh, &media_payload).await;
+
+    // Read the datagram on the peer side
+    let received = tokio::time::timeout(
+        Duration::from_secs(2),
+        peer_transport.connection().read_datagram(),
+    )
+    .await
+    .expect("should receive media within timeout")
+    .expect("read_datagram ok");
+
+    // Verify tagged format: [8-byte room_hash][media_payload]
+    assert!(received.len() >= 8);
+    let mut recv_hash = [0u8; 8];
+    recv_hash.copy_from_slice(&received[..8]);
+    assert_eq!(recv_hash, rh, "room hash must match");
+    assert_eq!(
+        &received[8..],
+        &media_payload[..],
+        "media payload must match"
+    );
+
+    drop(peer_transport);
+}
+
+// ───── 12. Multiple global rooms: each hashes independently ─────────
+
+#[tokio::test]
+async fn multiple_global_rooms_independent_hashes() {
+    let global: HashSet<String> = ["podcast", "lobby", "arena"]
+        .iter()
+        .map(|s| s.to_string())
+        .collect();
+    let fm = create_test_fm(global);
+
+    let hashes: Vec<[u8; 8]> = ["podcast", "lobby", "arena"]
+        .iter()
+        .map(|r| fm.global_room_hash(r))
+        .collect();
+
+    // All different
+    assert_ne!(hashes[0], hashes[1]);
+    assert_ne!(hashes[1], hashes[2]);
+    assert_ne!(hashes[0], hashes[2]);
+}
+
+// ───── 13. is_global_room edge cases ────────────────────────────────
+
+#[tokio::test]
+async fn is_global_room_exact_match_required_for_static() {
+    let global: HashSet<String> = ["podcast"].iter().map(|s| s.to_string()).collect();
+    let fm = create_test_fm(global);
+
+    // Substring/prefix should NOT match
+    assert!(!fm.is_global_room("podcast-extra"));
+    assert!(!fm.is_global_room("pod"));
+    assert!(!fm.is_global_room("podcastt"));
+}

crates/wzp-relay/tests/handshake_integration.rs

@@ -63,11 +63,11 @@ async fn handshake_succeeds() {
         accept_handshake(server_t.as_ref(), &callee_seed).await
     });
 
-    let caller_session = perform_handshake(client_transport.as_ref(), &caller_seed)
+    let caller_session = perform_handshake(client_transport.as_ref(), &caller_seed, None)
         .await
         .expect("perform_handshake should succeed");
 
-    let (callee_session, chosen_profile) = callee_handle
+    let (callee_session, chosen_profile, _caller_fp, _caller_alias) = callee_handle
         .await
         .expect("join callee task")
         .expect("accept_handshake should succeed");
@@ -124,11 +124,11 @@ async fn handshake_verifies_identity() {
         accept_handshake(server_t.as_ref(), &callee_seed).await
     });
 
-    let caller_session = perform_handshake(client_transport.as_ref(), &caller_seed)
+    let caller_session = perform_handshake(client_transport.as_ref(), &caller_seed, None)
         .await
         .expect("handshake must succeed even with different identities");
 
-    let (callee_session, _profile) = callee_handle
+    let (callee_session, _profile, _caller_fp, _caller_alias) = callee_handle
         .await
         .expect("join")
         .expect("accept_handshake must succeed");
@@ -183,7 +183,7 @@ async fn auth_then_handshake() {
         };
 
         // 2. Run the cryptographic handshake
-        let (session, profile) = accept_handshake(server_t.as_ref(), &callee_seed)
+        let (session, profile, _caller_fp, _caller_alias) = accept_handshake(server_t.as_ref(), &callee_seed)
             .await
             .expect("accept_handshake after auth");
 
@@ -199,7 +199,7 @@ async fn auth_then_handshake() {
         .await
         .expect("send AuthToken");
 
-    let caller_session = perform_handshake(client_transport.as_ref(), &caller_seed)
+    let caller_session = perform_handshake(client_transport.as_ref(), &caller_seed, None)
         .await
         .expect("perform_handshake after auth");
 
@@ -270,6 +270,7 @@ async fn handshake_rejects_bad_signature() {
         ephemeral_pub,
         signature,
         supported_profiles: vec![wzp_proto::QualityProfile::GOOD],
+        alias: None,
     };
 
     client_transport

crates/wzp-relay/tests/hole_punching.rs

@@ -0,0 +1,298 @@
+//! Phase 3 integration tests for hole-punching advertising
+//! (PRD: .taskmaster/docs/prd_hole_punching.txt).
+//!
+//! These verify the end-to-end protocol cross-wiring:
+//!   caller (places offer with caller_reflexive_addr=A)
+//!     → relay (stashes A in registry)
+//!       → callee (reads A off the forwarded offer)
+//!   callee (sends AcceptTrusted answer with callee_reflexive_addr=B)
+//!     → relay (stashes B, emits CallSetup to both parties)
+//!       → caller receives CallSetup.peer_direct_addr = B
+//!       → callee receives CallSetup.peer_direct_addr = A
+//!
+//! The actual QUIC hole-punch race is a Phase 3.5 follow-up.
+//! These tests only cover the signal-plane plumbing — that the
+//! addrs make it from each peer's offer/answer through the relay
+//! cross-wiring back out in CallSetup with the peer's addr.
+//!
+//! We drive the call registry + a minimal routing function
+//! directly instead of spinning up a full relay process — easier
+//! to reason about, no real network, and what we actually want to
+//! test is the cross-wiring logic, not the whole signal stack.
+
+use wzp_proto::{CallAcceptMode, SignalMessage};
+use wzp_relay::call_registry::CallRegistry;
+
+/// Helper: simulate the relay's handling of a DirectCallOffer. In
+/// `wzp-relay/src/main.rs` this is the match arm that creates the
+/// call in the registry and stashes the caller's reflex addr.
+fn handle_offer(reg: &mut CallRegistry, offer: &SignalMessage) -> String {
+    match offer {
+        SignalMessage::DirectCallOffer {
+            caller_fingerprint,
+            target_fingerprint,
+            call_id,
+            caller_reflexive_addr,
+            ..
+        } => {
+            reg.create_call(
+                call_id.clone(),
+                caller_fingerprint.clone(),
+                target_fingerprint.clone(),
+            );
+            reg.set_caller_reflexive_addr(call_id, caller_reflexive_addr.clone());
+            call_id.clone()
+        }
+        _ => panic!("not an offer"),
+    }
+}
+
+/// Helper: simulate the relay's handling of a DirectCallAnswer +
+/// the subsequent CallSetup emission. Returns the two CallSetup
+/// messages the relay would push: (for_caller, for_callee).
+fn handle_answer_and_build_setups(
+    reg: &mut CallRegistry,
+    answer: &SignalMessage,
+) -> (SignalMessage, SignalMessage) {
+    let (call_id, mode, callee_addr) = match answer {
+        SignalMessage::DirectCallAnswer {
+            call_id,
+            accept_mode,
+            callee_reflexive_addr,
+            ..
+        } => (call_id.clone(), *accept_mode, callee_reflexive_addr.clone()),
+        _ => panic!("not an answer"),
+    };
+
+    reg.set_callee_reflexive_addr(&call_id, callee_addr);
+    let room = format!("call-{call_id}");
+    reg.set_active(&call_id, mode, room.clone());
+
+    let (caller_addr, callee_addr) = {
+        let c = reg.get(&call_id).unwrap();
+        (
+            c.caller_reflexive_addr.clone(),
+            c.callee_reflexive_addr.clone(),
+        )
+    };
+
+    let setup_for_caller = SignalMessage::CallSetup {
+        call_id: call_id.clone(),
+        room: room.clone(),
+        relay_addr: "203.0.113.5:4433".into(),
+        peer_direct_addr: callee_addr,
+        peer_local_addrs: Vec::new(),
+        peer_mapped_addr: None,
+    };
+    let setup_for_callee = SignalMessage::CallSetup {
+        call_id,
+        room,
+        relay_addr: "203.0.113.5:4433".into(),
+        peer_direct_addr: caller_addr,
+        peer_local_addrs: Vec::new(),
+        peer_mapped_addr: None,
+    };
+    (setup_for_caller, setup_for_callee)
+}
+
+fn mk_offer(call_id: &str, caller_reflexive_addr: Option<&str>) -> SignalMessage {
+    SignalMessage::DirectCallOffer {
+        caller_fingerprint: "alice".into(),
+        caller_alias: None,
+        target_fingerprint: "bob".into(),
+        call_id: call_id.into(),
+        identity_pub: [0; 32],
+        ephemeral_pub: [0; 32],
+        signature: vec![],
+        supported_profiles: vec![],
+        caller_reflexive_addr: caller_reflexive_addr.map(String::from),
+        caller_local_addrs: Vec::new(),
+        caller_mapped_addr: None,
+        caller_build_version: None,
+    }
+}
+
+fn mk_answer(
+    call_id: &str,
+    mode: CallAcceptMode,
+    callee_reflexive_addr: Option<&str>,
+) -> SignalMessage {
+    SignalMessage::DirectCallAnswer {
+        call_id: call_id.into(),
+        accept_mode: mode,
+        identity_pub: None,
+        ephemeral_pub: None,
+        signature: None,
+        chosen_profile: None,
+        callee_reflexive_addr: callee_reflexive_addr.map(String::from),
+        callee_local_addrs: Vec::new(),
+        callee_mapped_addr: None,
+        callee_build_version: None,
+    }
+}
+
+// -----------------------------------------------------------------------
+// Test 1: both peers advertise — CallSetup cross-wires correctly
+// -----------------------------------------------------------------------
+
+#[test]
+fn both_peers_advertise_reflex_addrs_cross_wire_in_setup() {
+    let mut reg = CallRegistry::new();
+
+    let caller_addr = "192.0.2.1:4433";
+    let callee_addr = "198.51.100.9:4433";
+
+    let offer = mk_offer("c1", Some(caller_addr));
+    let call_id = handle_offer(&mut reg, &offer);
+    assert_eq!(call_id, "c1");
+    assert_eq!(
+        reg.get("c1").unwrap().caller_reflexive_addr.as_deref(),
+        Some(caller_addr)
+    );
+
+    let answer = mk_answer("c1", CallAcceptMode::AcceptTrusted, Some(callee_addr));
+    let (setup_caller, setup_callee) =
+        handle_answer_and_build_setups(&mut reg, &answer);
+
+    // The CALLER's setup should carry the CALLEE's addr as peer_direct_addr.
+    match setup_caller {
+        SignalMessage::CallSetup { peer_direct_addr, .. } => {
+            assert_eq!(
+                peer_direct_addr.as_deref(),
+                Some(callee_addr),
+                "caller's CallSetup must contain callee's addr"
+            );
+        }
+        _ => panic!("wrong variant"),
+    }
+
+    // The CALLEE's setup should carry the CALLER's addr.
+    match setup_callee {
+        SignalMessage::CallSetup { peer_direct_addr, .. } => {
+            assert_eq!(
+                peer_direct_addr.as_deref(),
+                Some(caller_addr),
+                "callee's CallSetup must contain caller's addr"
+            );
+        }
+        _ => panic!("wrong variant"),
+    }
+}
+
+// -----------------------------------------------------------------------
+// Test 2: callee uses AcceptGeneric (privacy) — no addr leaks
+// -----------------------------------------------------------------------
+
+#[test]
+fn privacy_mode_answer_omits_callee_addr_from_setup() {
+    let mut reg = CallRegistry::new();
+    let caller_addr = "192.0.2.1:4433";
+
+    handle_offer(&mut reg, &mk_offer("c2", Some(caller_addr)));
+
+    // AcceptGeneric explicitly passes None for callee_reflexive_addr —
+    // the whole point is to hide the callee's IP from the caller.
+    let answer = mk_answer("c2", CallAcceptMode::AcceptGeneric, None);
+    let (setup_caller, setup_callee) =
+        handle_answer_and_build_setups(&mut reg, &answer);
+
+    // CALLER should see peer_direct_addr = None (privacy preserved).
+    match setup_caller {
+        SignalMessage::CallSetup { peer_direct_addr, .. } => {
+            assert!(
+                peer_direct_addr.is_none(),
+                "privacy mode must not leak callee addr to caller"
+            );
+        }
+        _ => panic!("wrong variant"),
+    }
+
+    // CALLEE still gets the caller's addr — only the callee opted for
+    // privacy, the caller already volunteered its addr in the offer.
+    match setup_callee {
+        SignalMessage::CallSetup { peer_direct_addr, .. } => {
+            assert_eq!(
+                peer_direct_addr.as_deref(),
+                Some(caller_addr),
+                "callee's CallSetup should still carry caller's volunteered addr"
+            );
+        }
+        _ => panic!("wrong variant"),
+    }
+}
+
+// -----------------------------------------------------------------------
+// Test 3: old caller (no addr) + new callee — relay path only
+// -----------------------------------------------------------------------
+
+#[test]
+fn pre_phase3_caller_leaves_both_setups_relay_only() {
+    let mut reg = CallRegistry::new();
+
+    // Pre-Phase-3 client doesn't know about caller_reflexive_addr
+    // so the field is None.
+    handle_offer(&mut reg, &mk_offer("c3", None));
+
+    // New callee advertises its addr — doesn't matter because
+    // without caller_reflexive_addr the caller has nothing to
+    // attempt a direct handshake to, so the cross-wiring should
+    // still leave the caller's CallSetup without peer_direct_addr.
+    let answer = mk_answer(
+        "c3",
+        CallAcceptMode::AcceptTrusted,
+        Some("198.51.100.9:4433"),
+    );
+    let (setup_caller, setup_callee) =
+        handle_answer_and_build_setups(&mut reg, &answer);
+
+    match setup_caller {
+        SignalMessage::CallSetup { peer_direct_addr, .. } => {
+            // Phase 3 relay behavior: we always inject whatever
+            // addrs are in the registry, regardless of who
+            // advertised. The caller here gets the callee's addr
+            // because the callee did advertise.
+            assert_eq!(peer_direct_addr.as_deref(), Some("198.51.100.9:4433"));
+        }
+        _ => panic!("wrong variant"),
+    }
+
+    // The callee's setup has no caller addr (pre-Phase-3 offer).
+    match setup_callee {
+        SignalMessage::CallSetup { peer_direct_addr, .. } => {
+            assert!(
+                peer_direct_addr.is_none(),
+                "callee should see no caller addr when offer was pre-Phase-3"
+            );
+        }
+        _ => panic!("wrong variant"),
+    }
+}
+
+// -----------------------------------------------------------------------
+// Test 4: neither side advertises — both CallSetups fall back cleanly
+// -----------------------------------------------------------------------
+
+#[test]
+fn neither_peer_advertises_both_setups_are_relay_only() {
+    let mut reg = CallRegistry::new();
+
+    handle_offer(&mut reg, &mk_offer("c4", None));
+    let answer = mk_answer("c4", CallAcceptMode::AcceptTrusted, None);
+    let (setup_caller, setup_callee) =
+        handle_answer_and_build_setups(&mut reg, &answer);
+
+    for (label, setup) in [("caller", setup_caller), ("callee", setup_callee)] {
+        match setup {
+            SignalMessage::CallSetup { peer_direct_addr, relay_addr, .. } => {
+                assert!(
+                    peer_direct_addr.is_none(),
+                    "{label}'s CallSetup must have no peer_direct_addr"
+                );
+                // Relay addr is always filled — that's the fallback
+                // path and the existing behavior.
+                assert!(!relay_addr.is_empty(), "{label} relay_addr must be set");
+            }
+            _ => panic!("wrong variant"),
+        }
+    }
+}

crates/wzp-relay/tests/multi_reflect.rs

@@ -0,0 +1,231 @@
+//! Phase 2 integration tests for multi-relay NAT reflection
+//! (PRD: .taskmaster/docs/prd_multi_relay_reflect.txt).
+//!
+//! These spin up one or two mock relays that implement the full
+//! pre-reflect dance — RegisterPresence → RegisterPresenceAck →
+//! Reflect → ReflectResponse — which is what the transient
+//! probe helper in `wzp_client::reflect::probe_reflect_addr` does
+//! against a real relay.
+//!
+//! Test matrix:
+//!   1. `probe_reflect_addr_happy_path`
+//!      — single mock relay, assert the probe helper returns the
+//!        observed addr as 127.0.0.1:<client ephemeral port>
+//!   2. `detect_nat_type_two_loopback_relays_is_cone`
+//!      — two mock relays, one client; loopback single-host means
+//!        every probe sees the same (127.0.0.1, same_port) so the
+//!        classifier returns `Cone` + a consensus addr
+//!   3. `detect_nat_type_dead_relay_is_unknown`
+//!      — one alive relay + one dead address; aggregator returns
+//!        `Unknown` with a non-empty `error` field on the failed
+//!        probe
+
+use std::net::{Ipv4Addr, SocketAddr};
+use std::sync::Arc;
+use std::time::Duration;
+
+use wzp_client::reflect::{detect_nat_type, probe_reflect_addr, NatType};
+use wzp_proto::{MediaTransport, SignalMessage};
+use wzp_transport::{create_endpoint, server_config, QuinnTransport};
+
+/// Minimal mock relay that loops accepting connections, handles
+/// RegisterPresence + Reflect, and responds correctly. Mirrors the
+/// two match arms from `wzp-relay/src/main.rs` that matter here.
+///
+/// Each accepted connection gets its own inner task so multiple
+/// simultaneous probes work.
+async fn spawn_mock_relay() -> (SocketAddr, tokio::task::JoinHandle<()>) {
+    let _ = rustls::crypto::ring::default_provider().install_default();
+    let (sc, _cert_der) = server_config();
+    let bind: SocketAddr = (Ipv4Addr::LOCALHOST, 0).into();
+    let endpoint = create_endpoint(bind, Some(sc)).expect("server endpoint");
+    let listen_addr = endpoint.local_addr().expect("local_addr");
+
+    let handle = tokio::spawn(async move {
+        loop {
+            // Accept the next incoming connection. `wzp_transport::accept`
+            // returns the established `quinn::Connection`.
+            let conn = match wzp_transport::accept(&endpoint).await {
+                Ok(c) => c,
+                Err(_) => break, // endpoint closed
+            };
+            let observed_addr = conn.remote_address();
+            let transport = Arc::new(QuinnTransport::new(conn));
+
+            // Per-connection handler. Keep servicing messages until
+            // the peer closes so one probe connection can do
+            // RegisterPresence → Ack → Reflect → Response without
+            // racing other incoming connections.
+            let t = transport;
+            tokio::spawn(async move {
+                loop {
+                    match t.recv_signal().await {
+                        Ok(Some(SignalMessage::RegisterPresence { .. })) => {
+                            let _ = t
+                                .send_signal(&SignalMessage::RegisterPresenceAck {
+                                    success: true,
+                                    error: None,
+                                    relay_build: None,
+                                    relay_region: None,
+                                    available_relays: Vec::new(),
+                                })
+                                .await;
+                        }
+                        Ok(Some(SignalMessage::Reflect)) => {
+                            let _ = t
+                                .send_signal(&SignalMessage::ReflectResponse {
+                                    observed_addr: observed_addr.to_string(),
+                                })
+                                .await;
+                        }
+                        Ok(Some(_other)) => { /* ignore */ }
+                        Ok(None) => break,
+                        Err(_) => break,
+                    }
+                }
+            });
+        }
+    });
+
+    (listen_addr, handle)
+}
+
+// -----------------------------------------------------------------------
+// Test 1: probe_reflect_addr against a single mock relay
+// -----------------------------------------------------------------------
+
+#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
+async fn probe_reflect_addr_happy_path() {
+    let (relay_addr, _relay_handle) = spawn_mock_relay().await;
+
+    let (observed, latency_ms) = tokio::time::timeout(
+        Duration::from_secs(3),
+        probe_reflect_addr(relay_addr, 2000, None),
+    )
+    .await
+    .expect("probe must complete within 3s")
+    .expect("probe must succeed");
+
+    assert_eq!(
+        observed.ip().to_string(),
+        "127.0.0.1",
+        "loopback test should see 127.0.0.1"
+    );
+    assert_ne!(observed.port(), 0, "observed port must be non-zero");
+    // Latency on same host is dominated by the handshake — generously
+    // allow up to 2s (the timeout) rather than picking a tight number
+    // that would be flaky on busy CI runners.
+    assert!(latency_ms < 2000, "latency {latency_ms}ms too high");
+}
+
+// -----------------------------------------------------------------------
+// Test 2: two loopback relays → probes succeed, classification is Unknown
+// -----------------------------------------------------------------------
+//
+// With the private-IP filter added in the NAT classifier, loopback
+// reflex addrs (127.0.0.1) are dropped before classification —
+// they can't possibly indicate public-internet NAT state. So the
+// test now asserts:
+//   - both probes succeed end-to-end (wire plumbing works)
+//   - both return 127.0.0.1 (same-host is visible)
+//   - the aggregated verdict is Unknown (no public probes)
+
+#[tokio::test(flavor = "multi_thread", worker_threads = 4)]
+async fn detect_nat_type_two_loopback_relays_probes_work_but_classify_unknown() {
+    let (addr_a, _h_a) = spawn_mock_relay().await;
+    let (addr_b, _h_b) = spawn_mock_relay().await;
+
+    let detection = detect_nat_type(
+        vec![
+            ("RelayA".into(), addr_a),
+            ("RelayB".into(), addr_b),
+        ],
+        2000,
+        None,
+    )
+    .await;
+
+    assert_eq!(detection.probes.len(), 2);
+    for p in &detection.probes {
+        assert!(
+            p.observed_addr.is_some(),
+            "probe {:?} failed: {:?}",
+            p.relay_name,
+            p.error
+        );
+    }
+    let observed_ips: Vec<String> = detection
+        .probes
+        .iter()
+        .map(|p| {
+            p.observed_addr
+                .as_ref()
+                .and_then(|s| s.parse::<SocketAddr>().ok())
+                .map(|a| a.ip().to_string())
+                .unwrap_or_default()
+        })
+        .collect();
+    assert_eq!(observed_ips[0], "127.0.0.1");
+    assert_eq!(observed_ips[1], "127.0.0.1");
+
+    // Classification: loopback probes are filtered out of the
+    // public-NAT classifier, so with 0 public probes the result
+    // is Unknown.
+    assert_eq!(
+        detection.nat_type,
+        NatType::Unknown,
+        "loopback-only probes must not contribute to public NAT classification"
+    );
+    assert!(detection.consensus_addr.is_none());
+}
+
+// -----------------------------------------------------------------------
+// Test 3: one alive relay + one dead address → Unknown
+// -----------------------------------------------------------------------
+
+#[tokio::test(flavor = "multi_thread", worker_threads = 4)]
+async fn detect_nat_type_dead_relay_is_unknown() {
+    let (alive_addr, _alive_handle) = spawn_mock_relay().await;
+
+    // Dead relay: a port that nothing is listening on. OS will drop
+    // the packets, the probe should time out within the 600ms budget
+    // we give it. Pick a port unlikely to be in use — port 1 on
+    // loopback works on every OS I care about and fails fast.
+    let dead_addr: SocketAddr = "127.0.0.1:1".parse().unwrap();
+
+    let detection = detect_nat_type(
+        vec![
+            ("Alive".into(), alive_addr),
+            ("Dead".into(), dead_addr),
+        ],
+        600, // tight timeout so the dead probe fails fast
+        None,
+    )
+    .await;
+
+    assert_eq!(detection.probes.len(), 2);
+
+    // Find the alive and dead probes by name (order of JoinSet
+    // completions is not guaranteed).
+    let alive = detection.probes.iter().find(|p| p.relay_name == "Alive").unwrap();
+    let dead = detection.probes.iter().find(|p| p.relay_name == "Dead").unwrap();
+
+    assert!(
+        alive.observed_addr.is_some(),
+        "alive probe must succeed: {:?}",
+        alive.error
+    );
+    assert!(
+        dead.observed_addr.is_none(),
+        "dead probe must fail, got addr {:?}",
+        dead.observed_addr
+    );
+    assert!(
+        dead.error.is_some(),
+        "dead probe must surface an error string"
+    );
+
+    // With only 1 successful probe, the classifier returns Unknown.
+    assert_eq!(detection.nat_type, NatType::Unknown);
+    assert!(detection.consensus_addr.is_none());
+}

crates/wzp-relay/tests/reflect.rs

@@ -0,0 +1,318 @@
+//! Integration tests for the "STUN for QUIC" reflect protocol
+//! (PRD: .taskmaster/docs/prd_reflect_over_quic.txt, Phase 1).
+//!
+//! We don't spin up the full relay binary — instead we exercise the
+//! same wire-level request/response dance with a mock relay loop
+//! that implements exactly the match arm added to
+//! `wzp-relay/src/main.rs`. This isolates the protocol test from the
+//! rest of the relay state (rooms, federation, call registry, ...).
+//!
+//! Three test cases:
+//!  1. `reflect_happy_path` — client sends `Reflect`, mock relay
+//!     replies with `ReflectResponse { observed_addr }`, client
+//!     parses it back to a `SocketAddr` and confirms the IP is
+//!     `127.0.0.1` and the port matches its own bound port.
+//!  2. `reflect_two_clients_distinct_ports` — two simultaneous
+//!     client connections on different ephemeral ports get back
+//!     different reflected ports, proving the relay uses
+//!     per-connection `remote_address` rather than a global.
+//!  3. `reflect_old_relay_times_out` — mock relay that *doesn't*
+//!     handle `Reflect`; client side times out in the expected
+//!     window and does not hang.
+//!
+//! The third test uses a `tokio::time::timeout` wrapper directly
+//! (the client-side `request_reflect` helper lives in
+//! `desktop/src-tauri/src/lib.rs` which isn't a library we can
+//! depend on from here, so we reproduce the timeout semantics
+//! inline).
+
+use std::net::{Ipv4Addr, SocketAddr};
+use std::sync::Arc;
+use std::time::Duration;
+
+use wzp_proto::{MediaTransport, SignalMessage};
+use wzp_transport::{client_config, create_endpoint, server_config, QuinnTransport};
+
+/// Spawn a minimal mock relay that loops over `recv_signal`,
+/// matches on `Reflect`, and responds with `ReflectResponse` using
+/// the remote_address observed for this connection. Mirrors the
+/// match arm in `crates/wzp-relay/src/main.rs`.
+async fn spawn_mock_relay_with_reflect(
+    server_transport: Arc<QuinnTransport>,
+) -> tokio::task::JoinHandle<()> {
+    tokio::spawn(async move {
+        // Observed remote address at the time the connection was
+        // accepted. Stable for the life of the connection under quinn's
+        // normal operation. This is exactly what the real relay does.
+        let observed = server_transport.connection().remote_address();
+        loop {
+            match server_transport.recv_signal().await {
+                Ok(Some(SignalMessage::Reflect)) => {
+                    let resp = SignalMessage::ReflectResponse {
+                        observed_addr: observed.to_string(),
+                    };
+                    // If the send fails the client has gone; just exit.
+                    if server_transport.send_signal(&resp).await.is_err() {
+                        break;
+                    }
+                }
+                Ok(Some(_other)) => {
+                    // Ignore anything else — not relevant to this test.
+                }
+                Ok(None) => break,
+                Err(_e) => break,
+            }
+        }
+    })
+}
+
+/// Spawn a mock relay that intentionally DOES NOT handle Reflect.
+/// Models a pre-Phase-1 relay — it keeps reading signal messages and
+/// logs them to stderr, but never produces a `ReflectResponse`.
+async fn spawn_mock_relay_without_reflect(
+    server_transport: Arc<QuinnTransport>,
+) -> tokio::task::JoinHandle<()> {
+    tokio::spawn(async move {
+        loop {
+            match server_transport.recv_signal().await {
+                Ok(Some(_msg)) => {
+                    // Deliberately do nothing. Old relay.
+                }
+                Ok(None) => break,
+                Err(_) => break,
+            }
+        }
+    })
+}
+
+/// Build an in-process QUIC client/server pair on loopback and
+/// return (client_transport, server_transport, endpoints). The
+/// endpoints tuple must be kept alive for the test duration.
+///
+/// `client_port_hint` of 0 means "let OS pick". Pass an explicit
+/// port to pin the client's source port (useful for the
+/// distinct-ports test).
+async fn connected_pair_with_port(
+    _client_port_hint: u16,
+) -> (Arc<QuinnTransport>, Arc<QuinnTransport>, (quinn::Endpoint, quinn::Endpoint)) {
+    let _ = rustls::crypto::ring::default_provider().install_default();
+
+    let (sc, _cert_der) = server_config();
+    let server_addr: SocketAddr = (Ipv4Addr::LOCALHOST, 0).into();
+    let server_ep = create_endpoint(server_addr, Some(sc)).expect("server endpoint");
+    let server_listen = server_ep.local_addr().expect("server local addr");
+
+    // Always bind the client to an ephemeral port — we'll read back
+    // the actual assigned port via `local_addr()` in the assertions.
+    let client_bind: SocketAddr = (Ipv4Addr::LOCALHOST, 0).into();
+    let client_ep = create_endpoint(client_bind, None).expect("client endpoint");
+
+    let server_ep_clone = server_ep.clone();
+    let accept_fut = tokio::spawn(async move {
+        let conn = wzp_transport::accept(&server_ep_clone).await.expect("accept");
+        Arc::new(QuinnTransport::new(conn))
+    });
+
+    let client_conn =
+        wzp_transport::connect(&client_ep, server_listen, "localhost", client_config())
+            .await
+            .expect("connect");
+    let client_transport = Arc::new(QuinnTransport::new(client_conn));
+    let server_transport = accept_fut.await.expect("join accept task");
+
+    (client_transport, server_transport, (server_ep, client_ep))
+}
+
+// -----------------------------------------------------------------------
+// Test 1: happy path — client learns its own port via Reflect
+// -----------------------------------------------------------------------
+
+#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
+async fn reflect_happy_path() {
+    let (client_transport, server_transport, (_server_ep, client_ep)) =
+        connected_pair_with_port(0).await;
+
+    // Grab the client's actual bound port so we can cross-check
+    // against the reflected response.
+    let client_port = client_ep
+        .local_addr()
+        .expect("client local addr")
+        .port();
+    assert_ne!(client_port, 0, "client must have a real bound port");
+
+    // Start the mock relay's reflect handler.
+    let _relay_handle = spawn_mock_relay_with_reflect(Arc::clone(&server_transport)).await;
+
+    // Client sends Reflect and awaits the response. The real
+    // request_reflect helper in desktop/src-tauri/src/lib.rs uses a
+    // oneshot channel driven off the spawned recv loop; here we just
+    // do it inline because there's no spawned loop yet in this test
+    // — this isolates the wire protocol from the client-side state
+    // machine.
+    client_transport
+        .send_signal(&SignalMessage::Reflect)
+        .await
+        .expect("send Reflect");
+
+    let resp = tokio::time::timeout(Duration::from_secs(2), client_transport.recv_signal())
+        .await
+        .expect("reflect response should arrive within 2s")
+        .expect("recv_signal ok")
+        .expect("some message");
+
+    let observed_addr = match resp {
+        SignalMessage::ReflectResponse { observed_addr } => observed_addr,
+        other => panic!("expected ReflectResponse, got {:?}", std::mem::discriminant(&other)),
+    };
+
+    let parsed: SocketAddr = observed_addr
+        .parse()
+        .expect("ReflectResponse.observed_addr must parse as SocketAddr");
+
+    // The relay should see the client on 127.0.0.1 (loopback in the
+    // test harness) and on the client's bound ephemeral port.
+    assert_eq!(parsed.ip().to_string(), "127.0.0.1");
+    assert_eq!(
+        parsed.port(),
+        client_port,
+        "reflected port must match the client's local_addr port"
+    );
+
+    drop(client_transport);
+    drop(server_transport);
+}
+
+// -----------------------------------------------------------------------
+// Test 2: two clients get DIFFERENT reflected ports
+// -----------------------------------------------------------------------
+
+#[tokio::test(flavor = "multi_thread", worker_threads = 4)]
+async fn reflect_two_clients_distinct_ports() {
+    let _ = rustls::crypto::ring::default_provider().install_default();
+
+    // Shared server: one endpoint, two incoming accepts.
+    let (sc, _cert_der) = server_config();
+    let server_addr: SocketAddr = (Ipv4Addr::LOCALHOST, 0).into();
+    let server_ep = create_endpoint(server_addr, Some(sc)).expect("server endpoint");
+    let server_listen = server_ep.local_addr().expect("server local addr");
+
+    // Accept two clients in parallel.
+    let server_ep_a = server_ep.clone();
+    let accept_a = tokio::spawn(async move {
+        let conn = wzp_transport::accept(&server_ep_a).await.expect("accept A");
+        Arc::new(QuinnTransport::new(conn))
+    });
+    let server_ep_b = server_ep.clone();
+    let accept_b = tokio::spawn(async move {
+        let conn = wzp_transport::accept(&server_ep_b).await.expect("accept B");
+        Arc::new(QuinnTransport::new(conn))
+    });
+
+    // Client A
+    let client_ep_a = create_endpoint((Ipv4Addr::LOCALHOST, 0).into(), None).expect("ep A");
+    let conn_a =
+        wzp_transport::connect(&client_ep_a, server_listen, "localhost", client_config())
+            .await
+            .expect("connect A");
+    let client_a = Arc::new(QuinnTransport::new(conn_a));
+    let port_a = client_ep_a.local_addr().unwrap().port();
+
+    // Client B
+    let client_ep_b = create_endpoint((Ipv4Addr::LOCALHOST, 0).into(), None).expect("ep B");
+    let conn_b =
+        wzp_transport::connect(&client_ep_b, server_listen, "localhost", client_config())
+            .await
+            .expect("connect B");
+    let client_b = Arc::new(QuinnTransport::new(conn_b));
+    let port_b = client_ep_b.local_addr().unwrap().port();
+
+    assert_ne!(
+        port_a, port_b,
+        "preconditions: OS must assign two clients different ephemeral ports"
+    );
+
+    let server_a = accept_a.await.expect("join A");
+    let server_b = accept_b.await.expect("join B");
+
+    // Spawn a reflect handler for each server-side transport.
+    let _relay_a = spawn_mock_relay_with_reflect(Arc::clone(&server_a)).await;
+    let _relay_b = spawn_mock_relay_with_reflect(Arc::clone(&server_b)).await;
+
+    // Each client requests reflect concurrently.
+    let reflect_for = |t: Arc<QuinnTransport>| async move {
+        t.send_signal(&SignalMessage::Reflect).await.expect("send");
+        let resp = tokio::time::timeout(Duration::from_secs(2), t.recv_signal())
+            .await
+            .expect("timeout")
+            .expect("ok")
+            .expect("some");
+        match resp {
+            SignalMessage::ReflectResponse { observed_addr } => observed_addr,
+            _ => panic!("wrong variant"),
+        }
+    };
+
+    let (addr_a, addr_b) = tokio::join!(reflect_for(client_a.clone()), reflect_for(client_b.clone()));
+
+    let parsed_a: SocketAddr = addr_a.parse().unwrap();
+    let parsed_b: SocketAddr = addr_b.parse().unwrap();
+
+    assert_eq!(parsed_a.port(), port_a, "client A's reflected port");
+    assert_eq!(parsed_b.port(), port_b, "client B's reflected port");
+    assert_ne!(
+        parsed_a.port(),
+        parsed_b.port(),
+        "each client must see its own port, not a shared one"
+    );
+
+    drop(client_a);
+    drop(client_b);
+    drop(server_a);
+    drop(server_b);
+}
+
+// -----------------------------------------------------------------------
+// Test 3: old relay never answers — client times out cleanly
+// -----------------------------------------------------------------------
+
+#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
+async fn reflect_old_relay_times_out() {
+    let (client_transport, server_transport, _endpoints) =
+        connected_pair_with_port(0).await;
+
+    // Mock relay that ignores Reflect — simulates a pre-Phase-1 build.
+    let _relay_handle =
+        spawn_mock_relay_without_reflect(Arc::clone(&server_transport)).await;
+
+    client_transport
+        .send_signal(&SignalMessage::Reflect)
+        .await
+        .expect("send Reflect");
+
+    // 1100ms ceiling matches the 1s timeout baked into
+    // get_reflected_address plus a tiny bit of slack. If this
+    // regression ever fires it probably means recv_signal blocked
+    // longer than expected and the Tauri command would hang the UI.
+    let start = std::time::Instant::now();
+    let result =
+        tokio::time::timeout(Duration::from_millis(1100), client_transport.recv_signal()).await;
+    let elapsed = start.elapsed();
+
+    assert!(
+        result.is_err(),
+        "recv_signal must time out when the relay ignores Reflect"
+    );
+    assert!(
+        elapsed >= Duration::from_millis(1000),
+        "timeout fired too early ({:?})",
+        elapsed
+    );
+    assert!(
+        elapsed < Duration::from_millis(1200),
+        "timeout fired too late ({:?}), client would feel unresponsive",
+        elapsed
+    );
+
+    drop(client_transport);
+    drop(server_transport);
+}

crates/wzp-transport/Cargo.toml

@@ -15,6 +15,7 @@ tracing = { workspace = true }
 async-trait = { workspace = true }
 serde_json = "1"
 rustls = { version = "0.23", default-features = false, features = ["ring", "std"] }
+socket2 = { workspace = true }
 rcgen = "0.13"
 ed25519-dalek = { workspace = true }
 hkdf = { workspace = true }

crates/wzp-transport/src/config.rs

@@ -123,7 +123,6 @@ fn transport_config() -> quinn::TransportConfig {
     config.keep_alive_interval(Some(Duration::from_secs(5)));
 
     // Enable DATAGRAM extension for unreliable media packets.
-    // Allow datagrams up to 1200 bytes (conservative for lossy links).
     config.datagram_receive_buffer_size(Some(65536));
 
     // Conservative flow control for bandwidth-constrained links
@@ -134,6 +133,26 @@ fn transport_config() -> quinn::TransportConfig {
     // Aggressive initial RTT estimate for high-latency links
     config.initial_rtt(Duration::from_millis(300));
 
+    // PMTUD (Path MTU Discovery) — quinn 0.11 enables this by default but
+    // with conservative bounds (initial 1200, upper 1452). We keep the safe
+    // initial_mtu of 1200 so the first packets always get through, but raise
+    // upper_bound so the binary search can discover larger MTUs on paths that
+    // support them. Typical results:
+    //   - Ethernet/fiber: discovers ~1452 (Ethernet MTU minus IP/UDP/QUIC)
+    //   - WireGuard/VPN: discovers ~1380-1420
+    //   - Starlink: discovers ~1400-1452
+    //   - Cellular: stays at 1200-1300
+    // Black hole detection automatically falls back to 1200 if probes fail.
+    // This matters for future video frames which can be 1-50 KB and benefit
+    // from fewer application-layer fragments per frame.
+    let mut mtu_config = quinn::MtuDiscoveryConfig::default();
+    mtu_config
+        .upper_bound(1452)
+        .interval(Duration::from_secs(300))       // re-probe every 5 min
+        .black_hole_cooldown(Duration::from_secs(30)); // retry faster on lossy links
+    config.mtu_discovery_config(Some(mtu_config));
+    config.initial_mtu(1200); // safe starting point
+
     config
 }
 

crates/wzp-transport/src/connection.rs

@@ -39,6 +39,71 @@ pub async fn connect(
     Ok(connection)
 }
 
+/// Create an IPv6-only QUIC endpoint with `IPV6_V6ONLY=1`.
+///
+/// Tries `[::]:preferred_port` first (same port as the IPv4 signal
+/// endpoint — allowed on Linux/Android when the AFs differ and
+/// V6ONLY is set). Falls back to `[::]:0` (OS-assigned) if the
+/// preferred port is already taken.
+///
+/// Must be called from within a tokio runtime (quinn needs the
+/// async runtime handle for its I/O driver).
+pub fn create_ipv6_endpoint(
+    preferred_port: u16,
+    server_config: Option<quinn::ServerConfig>,
+) -> Result<quinn::Endpoint, TransportError> {
+    use socket2::{Domain, Protocol, Socket, Type};
+    use std::net::{Ipv6Addr, SocketAddrV6};
+
+    let sock = Socket::new(Domain::IPV6, Type::DGRAM, Some(Protocol::UDP))
+        .map_err(|e| TransportError::Internal(format!("ipv6 socket: {e}")))?;
+
+    // Critical: IPv6-only so this socket never intercepts IPv4.
+    // On Android some kernels default to V6ONLY=1 anyway, but we
+    // set it explicitly for cross-platform consistency.
+    sock.set_only_v6(true)
+        .map_err(|e| TransportError::Internal(format!("set_only_v6: {e}")))?;
+
+    sock.set_reuse_address(true)
+        .map_err(|e| TransportError::Internal(format!("set_reuse_address: {e}")))?;
+
+    // Try the preferred port (same as IPv4 signal endpoint), fall
+    // back to ephemeral if the OS rejects it.
+    let bind_addr = SocketAddrV6::new(Ipv6Addr::UNSPECIFIED, preferred_port, 0, 0);
+    if let Err(e) = sock.bind(&bind_addr.into()) {
+        if preferred_port != 0 {
+            tracing::debug!(
+                preferred_port,
+                error = %e,
+                "ipv6 bind to preferred port failed, falling back to ephemeral"
+            );
+            let fallback = SocketAddrV6::new(Ipv6Addr::UNSPECIFIED, 0, 0, 0);
+            sock.bind(&fallback.into())
+                .map_err(|e| TransportError::Internal(format!("ipv6 bind fallback: {e}")))?;
+        } else {
+            return Err(TransportError::Internal(format!("ipv6 bind: {e}")));
+        }
+    }
+
+    sock.set_nonblocking(true)
+        .map_err(|e| TransportError::Internal(format!("set_nonblocking: {e}")))?;
+
+    let udp_socket: std::net::UdpSocket = sock.into();
+
+    let runtime = quinn::default_runtime()
+        .ok_or_else(|| TransportError::Internal("no async runtime for ipv6 endpoint".into()))?;
+
+    let endpoint = quinn::Endpoint::new(
+        quinn::EndpointConfig::default(),
+        server_config,
+        udp_socket,
+        runtime,
+    )
+    .map_err(|e| TransportError::Internal(format!("ipv6 endpoint: {e}")))?;
+
+    Ok(endpoint)
+}
+
 /// Accept the next incoming connection on an endpoint.
 pub async fn accept(endpoint: &quinn::Endpoint) -> Result<quinn::Connection, TransportError> {
     let incoming = endpoint

crates/wzp-transport/src/lib.rs

@@ -23,7 +23,12 @@ pub mod quic;
 pub mod reliable;
 
 pub use config::{client_config, server_config, server_config_from_seed, tls_fingerprint};
-pub use connection::{accept, connect, create_endpoint};
+pub use connection::{accept, connect, create_endpoint, create_ipv6_endpoint};
 pub use path_monitor::PathMonitor;
-pub use quic::QuinnTransport;
+pub use quic::{QuinnPathSnapshot, QuinnTransport};
 pub use wzp_proto::{MediaTransport, PathQuality, TransportError};
+
+// Re-export the quinn Endpoint type so downstream crates (wzp-desktop) can
+// thread a shared endpoint between signaling and media connections without
+// needing to depend on quinn directly.
+pub use quinn::Endpoint;

crates/wzp-transport/src/path_monitor.rs

@@ -2,11 +2,17 @@
 //!
 //! Tracks packet loss (via sequence number gaps), RTT, jitter, and bandwidth.
 
+use std::collections::VecDeque;
+
 use wzp_proto::PathQuality;
 
 /// EWMA smoothing factor.
 const ALPHA: f64 = 0.1;
 
+/// Maximum number of RTT samples in the jitter variance sliding window.
+/// At ~50 packets/sec (20 ms frame), 10 samples ≈ 200 ms.
+const JITTER_VARIANCE_WINDOW_SIZE: usize = 10;
+
 /// Monitors network path quality metrics.
 pub struct PathMonitor {
     /// EWMA-smoothed loss percentage (0.0 - 100.0).
@@ -31,6 +37,8 @@ pub struct PathMonitor {
     last_rtt_ms: Option<f64>,
     /// Whether we have any observations yet.
     initialized: bool,
+    /// Sliding window of recent RTT samples for variance calculation.
+    rtt_window: VecDeque<f64>,
 }
 
 impl PathMonitor {
@@ -51,6 +59,7 @@ impl PathMonitor {
             total_received: 0,
             last_rtt_ms: None,
             initialized: false,
+            rtt_window: VecDeque::with_capacity(JITTER_VARIANCE_WINDOW_SIZE),
         }
     }
 
@@ -122,6 +131,12 @@ impl PathMonitor {
         } else {
             self.rtt_ewma = ALPHA * rtt + (1.0 - ALPHA) * self.rtt_ewma;
         }
+
+        // Maintain sliding window for variance calculation
+        if self.rtt_window.len() >= JITTER_VARIANCE_WINDOW_SIZE {
+            self.rtt_window.pop_front();
+        }
+        self.rtt_window.push_back(rtt);
     }
 
     /// Get the current estimated path quality.
@@ -155,6 +170,20 @@ impl PathMonitor {
         0
     }
 
+    /// Compute the jitter (RTT standard deviation) over the sliding window.
+    ///
+    /// Returns the standard deviation in milliseconds, or 0.0 if insufficient
+    /// samples. Used by `DredTuner` for spike detection.
+    pub fn jitter_variance_ms(&self) -> f64 {
+        let n = self.rtt_window.len();
+        if n < 2 {
+            return 0.0;
+        }
+        let mean = self.rtt_window.iter().sum::<f64>() / n as f64;
+        let var = self.rtt_window.iter().map(|r| (r - mean).powi(2)).sum::<f64>() / n as f64;
+        var.sqrt()
+    }
+
     /// Detect whether a network handoff likely occurred.
     ///
     /// Returns `true` if the most recent RTT jitter measurement exceeds 3x

crates/wzp-transport/src/quic.rs

@@ -13,6 +13,29 @@ use crate::datagram;
 use crate::path_monitor::PathMonitor;
 use crate::reliable;
 
+/// Snapshot of quinn's QUIC-level path statistics.
+///
+/// Provides more accurate loss/RTT data than `PathMonitor`'s sequence-gap
+/// heuristic because quinn sees ACK frames and congestion signals directly.
+#[derive(Clone, Copy, Debug)]
+pub struct QuinnPathSnapshot {
+    /// Smoothed RTT in milliseconds (from quinn's congestion controller).
+    pub rtt_ms: u32,
+    /// Cumulative loss percentage (lost_packets / sent_packets × 100).
+    pub loss_pct: f32,
+    /// Total congestion events observed by the QUIC stack.
+    pub congestion_events: u64,
+    /// Current congestion window in bytes.
+    pub cwnd: u64,
+    /// Total packets sent on this path.
+    pub sent_packets: u64,
+    /// Total packets lost on this path.
+    pub lost_packets: u64,
+    /// Current PMTUD-discovered maximum datagram payload size (bytes).
+    /// Starts at `initial_mtu` (1200) and grows as PMTUD probes succeed.
+    pub current_mtu: usize,
+}
+
 /// QUIC-based transport implementing the `MediaTransport` trait.
 pub struct QuinnTransport {
     connection: quinn::Connection,
@@ -33,6 +56,11 @@ impl QuinnTransport {
         &self.connection
     }
 
+    /// Remote address of the peer on this connection.
+    pub fn remote_address(&self) -> std::net::SocketAddr {
+        self.connection.remote_address()
+    }
+
     /// Send raw bytes as a QUIC datagram (no MediaPacket framing).
     pub fn send_raw_datagram(&self, data: &[u8]) -> Result<(), TransportError> {
         self.connection
@@ -61,6 +89,31 @@ impl QuinnTransport {
         datagram::max_datagram_payload(&self.connection)
     }
 
+    /// Snapshot of QUIC-level path stats from quinn, useful for DRED tuning.
+    ///
+    /// Returns `(rtt_ms, loss_pct, congestion_events)` derived from quinn's
+    /// internal congestion controller — more accurate than our own sequence-gap
+    /// heuristic in `PathMonitor` because quinn sees ACK frames directly.
+    pub fn quinn_path_stats(&self) -> QuinnPathSnapshot {
+        let stats = self.connection.stats();
+        let rtt_ms = stats.path.rtt.as_millis() as u32;
+        let loss_pct = if stats.path.sent_packets > 0 {
+            (stats.path.lost_packets as f32 / stats.path.sent_packets as f32) * 100.0
+        } else {
+            0.0
+        };
+        let current_mtu = self.connection.max_datagram_size().unwrap_or(1200);
+        QuinnPathSnapshot {
+            rtt_ms,
+            loss_pct,
+            congestion_events: stats.path.congestion_events,
+            cwnd: stats.path.cwnd,
+            sent_packets: stats.path.sent_packets,
+            lost_packets: stats.path.lost_packets,
+            current_mtu,
+        }
+    }
+
     /// Send an encoded [`TrunkFrame`] as a single QUIC datagram.
     pub fn send_trunk(&self, frame: &TrunkFrame) -> Result<(), TransportError> {
         let data = frame.encode();

crates/wzp-transport/src/reliable.rs

@@ -53,6 +53,13 @@ pub async fn recv_signal(recv: &mut quinn::RecvStream) -> Result<SignalMessage,
         .await
         .map_err(|e| TransportError::Internal(format!("stream read payload error: {e}")))?;
 
-    serde_json::from_slice(&payload)
-        .map_err(|e| TransportError::Internal(format!("signal deserialize error: {e}")))
+    serde_json::from_slice(&payload).map_err(|e| {
+        // Distinguish serde failures from transport failures so the
+        // caller (relay main loop, client recv loop) can continue on
+        // unknown-variant / parse errors instead of tearing down the
+        // whole signal connection. Forward-compat: adding a new
+        // `SignalMessage` variant in one side must not break the
+        // other side's signal connection.
+        TransportError::Deserialize(format!("{e}"))
+    })
 }

crates/wzp-web/static/wasm/package.json

@@ -0,0 +1,16 @@
+{
+  "name": "wzp-wasm",
+  "type": "module",
+  "description": "WarzonePhone WASM bindings — FEC (RaptorQ) + crypto (ChaCha20-Poly1305, X25519)",
+  "version": "0.1.0",
+  "files": [
+    "wzp_wasm_bg.wasm",
+    "wzp_wasm.js",
+    "wzp_wasm.d.ts"
+  ],
+  "main": "wzp_wasm.js",
+  "types": "wzp_wasm.d.ts",
+  "sideEffects": [
+    "./snippets/*"
+  ]
+}

crates/wzp-web/static/wasm/wzp_wasm.d.ts

@@ -0,0 +1,169 @@
+/* tslint:disable */
+/* eslint-disable */
+
+/**
+ * Symmetric encryption session using ChaCha20-Poly1305.
+ *
+ * Mirrors `wzp-crypto::session::ChaChaSession` for WASM.  Nonce derivation
+ * and key setup are identical so WASM and native peers interoperate.
+ */
+export class WzpCryptoSession {
+    free(): void;
+    [Symbol.dispose](): void;
+    /**
+     * Decrypt a media payload with AAD.
+     *
+     * Returns plaintext on success, or throws on auth failure.
+     */
+    decrypt(header_aad: Uint8Array, ciphertext: Uint8Array): Uint8Array;
+    /**
+     * Encrypt a media payload with AAD (typically the 12-byte MediaHeader).
+     *
+     * Returns `ciphertext || poly1305_tag` (plaintext.len() + 16 bytes).
+     */
+    encrypt(header_aad: Uint8Array, plaintext: Uint8Array): Uint8Array;
+    /**
+     * Create from a 32-byte shared secret (output of `WzpKeyExchange.derive_shared_secret`).
+     */
+    constructor(shared_secret: Uint8Array);
+    /**
+     * Current receive sequence number (for diagnostics / UI stats).
+     */
+    recv_seq(): number;
+    /**
+     * Current send sequence number (for diagnostics / UI stats).
+     */
+    send_seq(): number;
+}
+
+export class WzpFecDecoder {
+    free(): void;
+    [Symbol.dispose](): void;
+    /**
+     * Feed a received symbol.
+     *
+     * Returns the decoded block (concatenated original frames, unpadded) if
+     * enough symbols have been received to recover the block, or `undefined`.
+     */
+    add_symbol(block_id: number, symbol_idx: number, _is_repair: boolean, data: Uint8Array): Uint8Array | undefined;
+    /**
+     * Create a new FEC decoder.
+     *
+     * * `block_size` — expected number of source symbols per block.
+     * * `symbol_size` — padded byte size of each symbol (must match encoder).
+     */
+    constructor(block_size: number, symbol_size: number);
+}
+
+export class WzpFecEncoder {
+    free(): void;
+    [Symbol.dispose](): void;
+    /**
+     * Add a source symbol (audio frame).
+     *
+     * Returns encoded packets (all source + repair) when the block is complete,
+     * or `undefined` if the block is still accumulating.
+     *
+     * Each returned packet carries the 3-byte header:
+     *   `[block_id][symbol_idx][is_repair]` followed by `symbol_size` bytes.
+     */
+    add_symbol(data: Uint8Array): Uint8Array | undefined;
+    /**
+     * Force-flush the current (possibly partial) block.
+     *
+     * Returns all source + repair symbols with headers, or empty vec if no
+     * symbols have been accumulated.
+     */
+    flush(): Uint8Array;
+    /**
+     * Create a new FEC encoder.
+     *
+     * * `block_size` — number of source symbols (audio frames) per FEC block.
+     * * `symbol_size` — padded byte size of each symbol (default 256).
+     */
+    constructor(block_size: number, symbol_size: number);
+}
+
+/**
+ * X25519 key exchange: generate ephemeral keypair and derive shared secret.
+ *
+ * Usage from JS:
+ * ```js
+ * const kx = new WzpKeyExchange();
+ * const ourPub = kx.public_key();         // Uint8Array(32)
+ * // ... send ourPub to peer, receive peerPub ...
+ * const secret = kx.derive_shared_secret(peerPub); // Uint8Array(32)
+ * const session = new WzpCryptoSession(secret);
+ * ```
+ */
+export class WzpKeyExchange {
+    free(): void;
+    [Symbol.dispose](): void;
+    /**
+     * Derive a 32-byte session key from the peer's public key.
+     *
+     * Raw DH output is expanded via HKDF-SHA256 with info="warzone-session-key",
+     * matching `wzp-crypto::handshake::WarzoneKeyExchange::derive_session`.
+     */
+    derive_shared_secret(peer_public: Uint8Array): Uint8Array;
+    /**
+     * Generate a new random X25519 keypair.
+     */
+    constructor();
+    /**
+     * Our public key (32 bytes).
+     */
+    public_key(): Uint8Array;
+}
+
+export type InitInput = RequestInfo | URL | Response | BufferSource | WebAssembly.Module;
+
+export interface InitOutput {
+    readonly memory: WebAssembly.Memory;
+    readonly __wbg_wzpcryptosession_free: (a: number, b: number) => void;
+    readonly __wbg_wzpfecdecoder_free: (a: number, b: number) => void;
+    readonly __wbg_wzpfecencoder_free: (a: number, b: number) => void;
+    readonly __wbg_wzpkeyexchange_free: (a: number, b: number) => void;
+    readonly wzpcryptosession_decrypt: (a: number, b: number, c: number, d: number, e: number) => [number, number, number, number];
+    readonly wzpcryptosession_encrypt: (a: number, b: number, c: number, d: number, e: number) => [number, number, number, number];
+    readonly wzpcryptosession_new: (a: number, b: number) => [number, number, number];
+    readonly wzpcryptosession_recv_seq: (a: number) => number;
+    readonly wzpcryptosession_send_seq: (a: number) => number;
+    readonly wzpfecdecoder_add_symbol: (a: number, b: number, c: number, d: number, e: number, f: number) => [number, number];
+    readonly wzpfecdecoder_new: (a: number, b: number) => number;
+    readonly wzpfecencoder_add_symbol: (a: number, b: number, c: number) => [number, number];
+    readonly wzpfecencoder_flush: (a: number) => [number, number];
+    readonly wzpfecencoder_new: (a: number, b: number) => number;
+    readonly wzpkeyexchange_derive_shared_secret: (a: number, b: number, c: number) => [number, number, number, number];
+    readonly wzpkeyexchange_new: () => number;
+    readonly wzpkeyexchange_public_key: (a: number) => [number, number];
+    readonly __wbindgen_exn_store: (a: number) => void;
+    readonly __externref_table_alloc: () => number;
+    readonly __wbindgen_externrefs: WebAssembly.Table;
+    readonly __wbindgen_malloc: (a: number, b: number) => number;
+    readonly __externref_table_dealloc: (a: number) => void;
+    readonly __wbindgen_free: (a: number, b: number, c: number) => void;
+    readonly __wbindgen_start: () => void;
+}
+
+export type SyncInitInput = BufferSource | WebAssembly.Module;
+
+/**
+ * Instantiates the given `module`, which can either be bytes or
+ * a precompiled `WebAssembly.Module`.
+ *
+ * @param {{ module: SyncInitInput }} module - Passing `SyncInitInput` directly is deprecated.
+ *
+ * @returns {InitOutput}
+ */
+export function initSync(module: { module: SyncInitInput } | SyncInitInput): InitOutput;
+
+/**
+ * If `module_or_path` is {RequestInfo} or {URL}, makes a request and
+ * for everything else, calls `WebAssembly.instantiate` directly.
+ *
+ * @param {{ module_or_path: InitInput | Promise<InitInput> }} module_or_path - Passing `InitInput` directly is deprecated.
+ *
+ * @returns {Promise<InitOutput>}
+ */
+export default function __wbg_init (module_or_path?: { module_or_path: InitInput | Promise<InitInput> } | InitInput | Promise<InitInput>): Promise<InitOutput>;

crates/wzp-web/static/wasm/wzp_wasm_bg.wasm.d.ts

@@ -0,0 +1,27 @@
+/* tslint:disable */
+/* eslint-disable */
+export const memory: WebAssembly.Memory;
+export const __wbg_wzpcryptosession_free: (a: number, b: number) => void;
+export const __wbg_wzpfecdecoder_free: (a: number, b: number) => void;
+export const __wbg_wzpfecencoder_free: (a: number, b: number) => void;
+export const __wbg_wzpkeyexchange_free: (a: number, b: number) => void;
+export const wzpcryptosession_decrypt: (a: number, b: number, c: number, d: number, e: number) => [number, number, number, number];
+export const wzpcryptosession_encrypt: (a: number, b: number, c: number, d: number, e: number) => [number, number, number, number];
+export const wzpcryptosession_new: (a: number, b: number) => [number, number, number];
+export const wzpcryptosession_recv_seq: (a: number) => number;
+export const wzpcryptosession_send_seq: (a: number) => number;
+export const wzpfecdecoder_add_symbol: (a: number, b: number, c: number, d: number, e: number, f: number) => [number, number];
+export const wzpfecdecoder_new: (a: number, b: number) => number;
+export const wzpfecencoder_add_symbol: (a: number, b: number, c: number) => [number, number];
+export const wzpfecencoder_flush: (a: number) => [number, number];
+export const wzpfecencoder_new: (a: number, b: number) => number;
+export const wzpkeyexchange_derive_shared_secret: (a: number, b: number, c: number) => [number, number, number, number];
+export const wzpkeyexchange_new: () => number;
+export const wzpkeyexchange_public_key: (a: number) => [number, number];
+export const __wbindgen_exn_store: (a: number) => void;
+export const __externref_table_alloc: () => number;
+export const __wbindgen_externrefs: WebAssembly.Table;
+export const __wbindgen_malloc: (a: number, b: number) => number;
+export const __externref_table_dealloc: (a: number) => void;
+export const __wbindgen_free: (a: number, b: number, c: number) => void;
+export const __wbindgen_start: () => void;

desktop/.gitignore

@@ -0,0 +1,2 @@
+node_modules/
+dist/

desktop/.vite/deps/_metadata.json

@@ -0,0 +1,8 @@
+{
+  "hash": "9046c0bf",
+  "configHash": "ef0fc96f",
+  "lockfileHash": "d66891b1",
+  "browserHash": "8171ed59",
+  "optimized": {},
+  "chunks": {}
+}

desktop/.vite/deps/package.json

@@ -0,0 +1,3 @@
+{
+  "type": "module"
+}

desktop/index.html

@@ -0,0 +1,234 @@
+<!DOCTYPE html>
+<html lang="en">
+  <head>
+    <meta charset="UTF-8" />
+    <meta
+      name="viewport"
+      content="width=device-width, initial-scale=1.0, maximum-scale=1.0, minimum-scale=1.0, user-scalable=no, viewport-fit=cover"
+    />
+    <title>WarzonePhone</title>
+    <link rel="stylesheet" href="/src/style.css" />
+  </head>
+  <body>
+    <div id="app">
+
+      <!-- ═══════════════════════════════════════════════════════
+           LOBBY — default view, auto-connects signal on launch
+           ═══════════════════════════════════════════════════════ -->
+      <div id="lobby-screen">
+        <header class="lobby-header">
+          <div class="lobby-title-row">
+            <h1>WarzonePhone</h1>
+            <button id="settings-btn" class="icon-btn" title="Settings">&#9881;</button>
+          </div>
+          <div class="lobby-status-row">
+            <span id="lobby-dot" class="dot"></span>
+            <span id="lobby-relay-label" class="lobby-relay">Connecting...</span>
+            <span id="lobby-room-label" class="lobby-room">general</span>
+          </div>
+          <div class="lobby-identity">
+            <span id="lobby-identicon"></span>
+            <span id="lobby-fp" class="fp-display"></span>
+          </div>
+        </header>
+
+        <!-- User list -->
+        <div class="lobby-users-section">
+          <div class="lobby-users-header">
+            <span>Online</span>
+            <span id="lobby-user-count" class="badge">0</span>
+          </div>
+          <div id="lobby-user-list" class="lobby-user-list">
+            <div class="lobby-empty">No one else is here yet</div>
+          </div>
+        </div>
+
+        <!-- Voice join FAB -->
+        <div class="lobby-fab-row">
+          <button id="join-voice-btn" class="fab" title="Join Voice Chat">
+            <span class="fab-icon">&#x1F3A7;</span>
+            <span class="fab-label">Join Voice</span>
+          </button>
+        </div>
+
+        <!-- Incoming call banner -->
+        <div id="incoming-call-banner" class="incoming-banner hidden">
+          <div class="incoming-info">
+            <span id="incoming-identicon" class="incoming-identicon"></span>
+            <div>
+              <div id="incoming-caller-name" class="incoming-name">Unknown</div>
+              <div class="incoming-subtitle">Incoming call...</div>
+            </div>
+          </div>
+          <div class="incoming-actions">
+            <button id="accept-call-btn" class="btn-accept">Accept</button>
+            <button id="reject-call-btn" class="btn-reject">Reject</button>
+          </div>
+        </div>
+      </div>
+
+      <!-- ═══════════════════════════════════════════════════════
+           IN-CALL — voice active (room or direct)
+           ═══════════════════════════════════════════════════════ -->
+      <div id="call-screen" class="hidden">
+        <div class="call-header">
+          <div class="call-header-row">
+            <button id="back-to-lobby-btn" class="icon-btn small" title="Back to lobby">&#x2190;</button>
+            <div id="room-name" class="room-name"></div>
+            <button id="settings-btn-call" class="icon-btn small" title="Settings">&#9881;</button>
+          </div>
+          <div class="call-meta">
+            <span id="call-status" class="status-dot"></span>
+            <span id="call-timer" class="call-timer">0:00</span>
+          </div>
+        </div>
+        <div class="level-meter">
+          <div id="level-bar" class="level-bar-fill"></div>
+        </div>
+        <!-- Direct-call phone layout -->
+        <div id="direct-call-view" class="direct-call-view hidden">
+          <div id="dc-identicon" class="dc-identicon"></div>
+          <div id="dc-name" class="dc-name">Unknown</div>
+          <div id="dc-fp" class="dc-fp"></div>
+          <div id="dc-badge" class="dc-badge">Connecting...</div>
+        </div>
+        <!-- Room participants -->
+        <div id="participants" class="participants"></div>
+        <div class="controls">
+          <button id="mic-btn" class="control-btn" title="Toggle Mic (m)">
+            <span class="icon" id="mic-icon">Mic</span>
+          </button>
+          <button id="hangup-btn" class="control-btn hangup" title="Hang Up (q)">
+            <span class="icon">End</span>
+          </button>
+          <button id="spk-btn" class="control-btn" title="Toggle Speaker (s)">
+            <span class="icon" id="spk-icon">Spk</span>
+          </button>
+        </div>
+        <div id="stats" class="stats"></div>
+      </div>
+
+      <!-- ═══════════════════════════════════════════════════════
+           USER CONTEXT MENU (tap on user in lobby)
+           ═══════════════════════════════════════════════════════ -->
+      <div id="user-context-menu" class="context-menu hidden">
+        <div class="context-header">
+          <span id="ctx-identicon" class="ctx-identicon"></span>
+          <div>
+            <div id="ctx-name" class="ctx-name">User</div>
+            <div id="ctx-fp" class="ctx-fp"></div>
+          </div>
+        </div>
+        <button id="ctx-call-btn" class="context-action">
+          <span>&#x1F4DE;</span> Direct Call
+        </button>
+        <button id="ctx-message-btn" class="context-action" disabled>
+          <span>&#x1F4AC;</span> Message (coming soon)
+        </button>
+        <button id="ctx-close-btn" class="context-action dim">Close</button>
+      </div>
+
+      <!-- ═══════════════════════════════════════════════════════
+           SETTINGS PANEL (overlay)
+           ═══════════════════════════════════════════════════════ -->
+      <div id="settings-panel" class="hidden">
+        <div class="settings-card">
+          <div class="settings-header">
+            <h2>Settings</h2>
+            <button id="settings-close" class="icon-btn">&times;</button>
+          </div>
+          <div class="settings-section">
+            <h3>Connection</h3>
+            <label>Default Room
+              <input id="s-room" type="text" />
+            </label>
+            <label>Alias
+              <input id="s-alias" type="text" />
+            </label>
+          </div>
+          <div class="settings-section">
+            <h3>Audio</h3>
+            <div class="quality-control">
+              <div class="quality-header">
+                <span class="setting-label">QUALITY</span>
+                <span id="s-quality-label" class="quality-value">Auto</span>
+              </div>
+              <input id="s-quality" type="range" min="0" max="6" step="1" value="6" />
+              <div class="quality-labels">
+                <span>Codec2 1.2k</span>
+                <span>Auto</span>
+              </div>
+            </div>
+            <label class="checkbox">
+              <input id="s-os-aec" type="checkbox" checked />
+              OS Echo Cancellation
+            </label>
+          </div>
+          <div class="settings-section">
+            <h3>Relays</h3>
+            <div id="s-relay-list"></div>
+            <div class="relay-add">
+              <input id="s-relay-name" type="text" placeholder="Name" style="flex:1" />
+              <input id="s-relay-addr" type="text" placeholder="host:port" style="flex:2" />
+              <button id="s-relay-add" class="secondary-btn small">Add</button>
+            </div>
+          </div>
+          <div class="settings-section">
+            <h3>Identity</h3>
+            <div>
+              <span class="setting-label">FINGERPRINT</span>
+              <div id="s-fingerprint" class="fp-display" style="margin-top:4px"></div>
+            </div>
+            <div style="margin-top:8px">
+              <span class="setting-label">IDENTITY FILE</span>
+              <div style="font-size:12px;opacity:0.6;margin-top:2px">~/.wzp/identity</div>
+            </div>
+          </div>
+          <div class="settings-section">
+            <h3>Network</h3>
+            <div>
+              <span class="setting-label">PUBLIC ADDRESS</span>
+              <span id="s-public-addr" style="color:var(--green);font-size:13px;margin-left:8px"></span>
+              <button id="s-reflect-btn" class="secondary-btn small" style="margin-left:8px">Detect</button>
+            </div>
+            <div style="margin-top:8px">
+              <button id="s-nat-detect-btn" class="secondary-btn" style="width:100%">Detect NAT</button>
+              <div id="s-nat-result" style="font-size:11px;margin-top:4px;opacity:0.7;white-space:pre-wrap"></div>
+            </div>
+          </div>
+          <div class="settings-section">
+            <h3>Debug</h3>
+            <label class="checkbox">
+              <input id="s-dred-debug" type="checkbox" />
+              DRED debug logs (verbose, dev only)
+            </label>
+            <label class="checkbox">
+              <input id="s-call-debug" type="checkbox" />
+              Call flow debug logs (trace every step of a call)
+            </label>
+            <label class="checkbox">
+              <input id="s-direct-only" type="checkbox" />
+              Direct-only mode (no relay fallback)
+            </label>
+            <label class="checkbox">
+              <input id="s-birthday-attack" type="checkbox" />
+              Birthday attack (extra ports for hard NAT — adds ~3s)
+            </label>
+          </div>
+          <div class="settings-section" id="s-call-debug-section" style="display:none">
+            <h3>Call Debug Log</h3>
+            <div id="s-call-debug-log" style="max-height:220px;overflow-y:auto;background:#0a0a0a;color:#e0e0e0;font-family:ui-monospace,Menlo,Monaco,'Courier New',monospace;font-size:10px;padding:6px;border-radius:4px;line-height:1.4;white-space:pre-wrap"></div>
+            <div style="display:flex;gap:6px;margin-top:6px">
+              <button id="s-call-debug-copy" class="secondary-btn" style="flex:1">Copy log</button>
+              <button id="s-call-debug-share" class="secondary-btn" style="flex:1">Share</button>
+              <button id="s-call-debug-clear" class="secondary-btn" style="flex:1">Clear log</button>
+            </div>
+            <small id="s-call-debug-copy-status" style="display:block;margin-top:4px;color:var(--text-dim);font-size:10px"></small>
+          </div>
+          <button id="settings-save" class="primary" style="margin-top:12px">Save</button>
+        </div>
+      </div>
+    </div>
+    <script type="module" src="/src/main.ts"></script>
+  </body>
+</html>

desktop/package.json

@@ -0,0 +1,19 @@
+{
+  "name": "wzp-desktop",
+  "private": true,
+  "version": "0.1.0",
+  "type": "module",
+  "scripts": {
+    "dev": "vite",
+    "build": "vite build",
+    "tauri": "tauri"
+  },
+  "dependencies": {
+    "@tauri-apps/api": "^2"
+  },
+  "devDependencies": {
+    "typescript": "^5",
+    "vite": "^6",
+    "@tauri-apps/cli": "^2"
+  }
+}

desktop/src-tauri/Cargo.toml

@@ -0,0 +1,108 @@
+[package]
+name = "wzp-desktop"
+version = "0.1.0"
+edition = "2024"
+description = "WarzonePhone Desktop — encrypted VoIP client"
+default-run = "wzp-desktop"
+
+# Library target — required for Tauri mobile (Android/iOS link the app as a cdylib)
+# and also used by the desktop binary below.
+#
+# `staticlib` was DROPPED from crate-type because rust-lang/rust#104707
+# documents that having staticlib alongside cdylib leaks non-exported
+# symbols from staticlibs into the cdylib. Bionic's private `__init_tcb`
+# / `pthread_create` symbols end up bound LOCALLY inside our .so instead
+# of resolved dynamically against libc.so at dlopen time — which crashes
+# at launch as soon as tao tries to std::thread::spawn() from the JNI
+# onCreate callback. The legacy wzp-android crate uses ["cdylib", "rlib"]
+# and runs fine on the same phone with the same NDK + Rust toolchain.
+#
+# iOS Tauri builds that actually need staticlib can re-add it behind a
+# target cfg if we ever ship on iOS.
+[lib]
+name = "wzp_desktop_lib"
+crate-type = ["cdylib", "rlib"]
+
+[[bin]]
+name = "wzp-desktop"
+path = "src/main.rs"
+
+[build-dependencies]
+tauri-build = { version = "2", features = [] }
+# cc is no longer needed — all C++ moved to crates/wzp-native (built with
+# cargo-ndk and loaded via libloading at runtime). wzp-desktop's .so on
+# Android is now pure Rust.
+
+[dependencies]
+tauri = { version = "2", features = [] }
+tauri-plugin-shell = "2"
+tauri-plugin-notification = "2"
+serde = { version = "1", features = ["derive"] }
+serde_json = "1"
+tokio = { version = "1", features = ["full"] }
+tracing = "0.1"
+tracing-subscriber = "0.3"
+anyhow = "1"
+rustls = { version = "0.23", default-features = false, features = ["ring", "std"] }
+
+# WarzonePhone crates — protocol layer is platform-independent
+wzp-proto = { path = "../../crates/wzp-proto" }
+wzp-codec = { path = "../../crates/wzp-codec" }
+wzp-fec = { path = "../../crates/wzp-fec" }
+wzp-crypto = { path = "../../crates/wzp-crypto" }
+wzp-transport = { path = "../../crates/wzp-transport" }
+
+# wzp-client pulls in CPAL on every desktop target and, additionally on
+# macOS, VoiceProcessingIO (coreaudio-rs behind the "vpio" feature). The
+# vpio feature MUST NOT be enabled on Windows / Linux because coreaudio-rs
+# is Apple-framework-only and will fail to build. Task #24 will add a
+# matching Windows Voice Capture DSP path behind its own feature; until
+# then, Windows desktops use plain CPAL with AEC disabled.
+
+# macOS: CPAL + VoiceProcessingIO (hardware AEC via Core Audio).
+[target.'cfg(target_os = "macos")'.dependencies]
+wzp-client = { path = "../../crates/wzp-client", features = ["audio", "vpio"] }
+
+# Windows: CPAL for playback + direct WASAPI for capture with OS-level
+# AEC (AudioCategory_Communications). The wzp-client `windows-aec`
+# feature swaps the default CPAL AudioCapture for a WASAPI one that
+# opens the mic under AudioCategory_Communications, turning on Windows's
+# communications audio processing chain (AEC, NS, AGC). The reference
+# signal for AEC is the system render mix, so echo from our CPAL
+# playback is cancelled automatically without extra plumbing.
+[target.'cfg(target_os = "windows")'.dependencies]
+wzp-client = { path = "../../crates/wzp-client", features = ["audio", "windows-aec"] }
+
+# Linux: CPAL playback+capture baseline. AEC is enabled via the top-level
+# `linux-aec` feature in wzp-desktop, which forwards to wzp-client/linux-aec.
+# Keeping it opt-in at the wzp-desktop level (rather than forcing it always
+# on here) lets `cargo tauri build` produce two variants from the same
+# source tree — a noAEC baseline and an AEC build — by toggling the feature
+# at build time: `cargo tauri build -- --features wzp-desktop/linux-aec`.
+[target.'cfg(target_os = "linux")'.dependencies]
+wzp-client = { path = "../../crates/wzp-client", features = ["audio"] }
+
+# Android: no CPAL, no vpio — audio goes through the standalone wzp-native
+# cdylib that we dlopen via libloading at runtime. See the wzp_native
+# module in src/.
+[target.'cfg(target_os = "android")'.dependencies]
+wzp-client = { path = "../../crates/wzp-client", default-features = false }
+# libloading: runtime dlopen of libwzp_native.so — the standalone cdylib
+# crate that owns all C++ (Oboe bridge). Keeps wzp-desktop's .so free of
+# any C/C++ static archives that would otherwise leak bionic's internal
+# pthread_create into our cdylib and trigger the __init_tcb crash.
+libloading = "0.8"
+# jni + ndk-context: called from android_audio.rs to invoke
+# AudioManager.setSpeakerphoneOn on the JVM side at runtime, so the
+# Oboe playout stream (opened with Usage::VoiceCommunication) can route
+# between earpiece and loud speaker without restarting.
+jni = "0.21"
+ndk-context = "0.1"
+
+[features]
+default = ["custom-protocol"]
+custom-protocol = ["tauri/custom-protocol"]
+# linux-aec: forwards to wzp-client/linux-aec so `cargo tauri build -- --features
+# wzp-desktop/linux-aec` enables the WebRTC AEC3 backend on Linux. No-op on
+# other targets because wzp-client/linux-aec is itself cfg(target_os = "linux").
+linux-aec = ["wzp-client/linux-aec"]

desktop/src-tauri/Info.plist

@@ -0,0 +1,21 @@
+<?xml version="1.0" encoding="UTF-8"?>
+<!DOCTYPE plist PUBLIC "-//Apple//DTD PLIST 1.0//EN" "http://www.apple.com/DTDs/PropertyList-1.0.dtd">
+<plist version="1.0">
+<dict>
+    <!--
+      Custom Info.plist keys merged into the bundled WarzonePhone.app by
+      tauri-bundler. The base Info.plist (CFBundleIdentifier, version,
+      etc.) is generated from tauri.conf.json — only put *additional*
+      keys here.
+
+      NSMicrophoneUsageDescription is required by macOS TCC for any
+      app that opens an audio input unit. Without this string the OS
+      silently denies CoreAudio capture (input callbacks return zeros)
+      and the app never appears in System Settings → Privacy &
+      Security → Microphone. This was the root cause of the desktop
+      mic regression where phones could not hear the desktop client.
+    -->
+    <key>NSMicrophoneUsageDescription</key>
+    <string>WarzonePhone needs microphone access to transmit your voice during calls.</string>
+</dict>
+</plist>

desktop/src-tauri/build.rs

@@ -0,0 +1,26 @@
+use std::process::Command;
+
+fn main() {
+    // Capture short git hash so the running app can prove which build it is.
+    // Falls back to "unknown" if git isn't available (e.g. when building from
+    // a tarball without a .git dir).
+    let git_hash = Command::new("git")
+        .args(["rev-parse", "--short", "HEAD"])
+        .output()
+        .ok()
+        .filter(|o| o.status.success())
+        .and_then(|o| String::from_utf8(o.stdout).ok())
+        .map(|s| s.trim().to_string())
+        .unwrap_or_else(|| "unknown".into());
+
+    println!("cargo:rustc-env=WZP_GIT_HASH={git_hash}");
+    println!("cargo:rerun-if-changed=../../.git/HEAD");
+    println!("cargo:rerun-if-changed=../../.git/refs/heads");
+
+    // No cc::Build of ANY kind on Android — all C++ lives in the standalone
+    // `wzp-native` crate which is built separately with cargo-ndk and loaded
+    // via libloading at runtime. See docs/incident-tauri-android-init-tcb.md
+    // for why this split exists.
+
+    tauri_build::build()
+}

desktop/src-tauri/capabilities/default.json

@@ -0,0 +1,30 @@
+{
+  "$schema": "../gen/schemas/desktop-schema.json",
+  "identifier": "default",
+  "description": "Default capability — grants core APIs (events, path, window, app, clipboard) to the main window on every platform we ship to.",
+  "windows": ["main"],
+  "platforms": [
+    "linux",
+    "macOS",
+    "windows",
+    "android",
+    "iOS"
+  ],
+  "permissions": [
+    "core:default",
+    "core:event:default",
+    "core:event:allow-listen",
+    "core:event:allow-unlisten",
+    "core:event:allow-emit",
+    "core:event:allow-emit-to",
+    "core:path:default",
+    "core:window:default",
+    "core:app:default",
+    "core:webview:default",
+    "shell:default",
+    "notification:default",
+    "notification:allow-notify",
+    "notification:allow-request-permission",
+    "notification:allow-is-permission-granted"
+  ]
+}

desktop/src-tauri/gen/android/app/src/main/AndroidManifest.xml

@@ -0,0 +1,40 @@
+<?xml version="1.0" encoding="utf-8"?>
+<manifest xmlns:android="http://schemas.android.com/apk/res/android">
+    <uses-permission android:name="android.permission.INTERNET" />
+    <uses-permission android:name="android.permission.RECORD_AUDIO" />
+    <uses-permission android:name="android.permission.MODIFY_AUDIO_SETTINGS" />
+    <uses-feature android:name="android.hardware.microphone" android:required="true" />
+
+    <!-- AndroidTV support -->
+    <uses-feature android:name="android.software.leanback" android:required="false" />
+
+    <application
+        android:icon="@mipmap/ic_launcher"
+        android:label="@string/app_name"
+        android:theme="@style/Theme.wzp_desktop"
+        android:usesCleartextTraffic="${usesCleartextTraffic}">
+        <activity
+            android:configChanges="orientation|keyboardHidden|keyboard|screenSize|locale|smallestScreenSize|screenLayout|uiMode"
+            android:launchMode="singleTask"
+            android:label="@string/main_activity_title"
+            android:name=".MainActivity"
+            android:exported="true">
+            <intent-filter>
+                <action android:name="android.intent.action.MAIN" />
+                <category android:name="android.intent.category.LAUNCHER" />
+                <!-- AndroidTV support -->
+                <category android:name="android.intent.category.LEANBACK_LAUNCHER" />
+            </intent-filter>
+        </activity>
+
+        <provider
+          android:name="androidx.core.content.FileProvider"
+          android:authorities="${applicationId}.fileprovider"
+          android:exported="false"
+          android:grantUriPermissions="true">
+          <meta-data
+            android:name="android.support.FILE_PROVIDER_PATHS"
+            android:resource="@xml/file_paths" />
+        </provider>
+    </application>
+</manifest>

desktop/src-tauri/gen/android/app/src/main/java/com/wzp/desktop/MainActivity.kt

@@ -0,0 +1,103 @@
+package com.wzp.desktop
+
+import android.Manifest
+import android.content.Context
+import android.content.pm.PackageManager
+import android.media.AudioManager
+import android.os.Bundle
+import android.util.Log
+import androidx.activity.enableEdgeToEdge
+import androidx.core.app.ActivityCompat
+import androidx.core.content.ContextCompat
+
+class MainActivity : TauriActivity() {
+  companion object {
+    private const val TAG = "WzpMainActivity"
+    private const val AUDIO_PERMISSIONS_REQUEST = 4242
+    private val REQUIRED_AUDIO_PERMISSIONS = arrayOf(
+      Manifest.permission.RECORD_AUDIO,
+      Manifest.permission.MODIFY_AUDIO_SETTINGS
+    )
+  }
+
+  override fun onCreate(savedInstanceState: Bundle?) {
+    enableEdgeToEdge()
+    super.onCreate(savedInstanceState)
+
+    // Request RECORD_AUDIO early so Oboe (inside libwzp_native.so) can open
+    // the AAudio input stream without silently failing. The grant is
+    // persisted, so after the first launch the dialog no longer appears.
+    // MODIFY_AUDIO_SETTINGS is needed to switch AudioManager mode + speaker.
+    val needsRequest = REQUIRED_AUDIO_PERMISSIONS.any {
+      ContextCompat.checkSelfPermission(this, it) != PackageManager.PERMISSION_GRANTED
+    }
+    if (needsRequest) {
+      Log.i(TAG, "requesting audio permissions")
+      ActivityCompat.requestPermissions(this, REQUIRED_AUDIO_PERMISSIONS, AUDIO_PERMISSIONS_REQUEST)
+    } else {
+      Log.i(TAG, "audio permissions already granted")
+      configureAudioForCall()
+    }
+  }
+
+  override fun onRequestPermissionsResult(
+    requestCode: Int,
+    permissions: Array<String>,
+    grantResults: IntArray
+  ) {
+    super.onRequestPermissionsResult(requestCode, permissions, grantResults)
+    if (requestCode == AUDIO_PERMISSIONS_REQUEST) {
+      val allGranted = grantResults.isNotEmpty() &&
+        grantResults.all { it == PackageManager.PERMISSION_GRANTED }
+      Log.i(TAG, "audio permissions result: allGranted=$allGranted grants=${grantResults.toList()}")
+      if (allGranted) {
+        configureAudioForCall()
+      }
+    }
+  }
+
+  /**
+   * Put the phone into VoIP call mode with handset (earpiece) as the
+   * default output. The Oboe playout stream is opened with
+   * Usage::VoiceCommunication which honours this routing, so:
+   *
+   *   MODE_IN_COMMUNICATION + speakerphoneOn=false  → earpiece (handset)
+   *   MODE_IN_COMMUNICATION + speakerphoneOn=true   → loudspeaker
+   *   MODE_IN_COMMUNICATION + bluetoothScoOn=true   → bluetooth headset
+   *
+   * The speaker/handset/BT toggle itself is wired up via the Tauri
+   * command `set_speakerphone(on)` in a follow-up build. For now the
+   * default is handset, matching the user's stated preference.
+   *
+   * STREAM_VOICE_CALL volume is cranked to max since the in-call volume
+   * slider is separate from media volume on most devices.
+   */
+  /**
+   * Pre-flight: only set volumes. Do NOT set MODE_IN_COMMUNICATION here —
+   * that hijacks the entire audio routing (music stops, BT A2DP drops to
+   * earpiece) even before a call starts. The Rust side sets the mode via
+   * JNI when the call engine actually starts, and restores MODE_NORMAL
+   * when the call ends.
+   */
+  private fun configureAudioForCall() {
+    try {
+      val am = getSystemService(Context.AUDIO_SERVICE) as AudioManager
+      Log.i(TAG, "audio state: mode=${am.mode} speaker=${am.isSpeakerphoneOn} " +
+        "voiceVol=${am.getStreamVolume(AudioManager.STREAM_VOICE_CALL)}/" +
+        "${am.getStreamMaxVolume(AudioManager.STREAM_VOICE_CALL)} " +
+        "musicVol=${am.getStreamVolume(AudioManager.STREAM_MUSIC)}/" +
+        "${am.getStreamMaxVolume(AudioManager.STREAM_MUSIC)}")
+
+      // Crank both voice-call and music volumes so nothing silent slips
+      // through regardless of which stream actually ends up driving.
+      val maxVoice = am.getStreamMaxVolume(AudioManager.STREAM_VOICE_CALL)
+      am.setStreamVolume(AudioManager.STREAM_VOICE_CALL, maxVoice, 0)
+      val maxMusic = am.getStreamMaxVolume(AudioManager.STREAM_MUSIC)
+      am.setStreamVolume(AudioManager.STREAM_MUSIC, maxMusic, 0)
+
+      Log.i(TAG, "volumes set: voiceVol=$maxVoice musicVol=$maxMusic (mode left at ${am.mode})")
+    } catch (e: Throwable) {
+      Log.e(TAG, "configureAudioForCall failed: ${e.message}", e)
+    }
+  }
+}

desktop/src-tauri/gen/schemas/capabilities.json

@@ -0,0 +1 @@
+{"default":{"identifier":"default","description":"Default capability — grants core APIs (events, path, window, app, clipboard) to the main window on every platform we ship to.","local":true,"windows":["main"],"permissions":["core:default","core:event:default","core:event:allow-listen","core:event:allow-unlisten","core:event:allow-emit","core:event:allow-emit-to","core:path:default","core:window:default","core:app:default","core:webview:default","shell:default","notification:default","notification:allow-notify","notification:allow-request-permission","notification:allow-is-permission-granted"],"platforms":["linux","macOS","windows","android","iOS"]}}

desktop/src-tauri/src/android_audio.rs

@@ -0,0 +1,359 @@
+//! Runtime bridge to Android's `AudioManager` for in-call audio routing.
+//!
+//! We own a quinn+Oboe VoIP pipeline entirely from Rust, but routing the
+//! playout stream between earpiece / loudspeaker / Bluetooth headset has to
+//! happen at the JVM level because those toggles are AudioManager-only.
+//! This module uses the global JavaVM handle that `ndk_context` exposes
+//! (populated by Tauri's mobile runtime) + the `jni` crate to reach into
+//! the Android framework without needing a Tauri plugin.
+//!
+//! All callers must be inside an Android target (`#[cfg(target_os = "android")]`).
+
+#![cfg(target_os = "android")]
+
+use jni::objects::{JObject, JString, JValue};
+use jni::JavaVM;
+
+/// Grab the JavaVM + current Activity from the ndk_context that Tauri's
+/// mobile runtime sets up at process startup.
+fn jvm_and_activity() -> Result<(JavaVM, JObject<'static>), String> {
+    let ctx = ndk_context::android_context();
+    let vm_ptr = ctx.vm() as *mut jni::sys::JavaVM;
+    if vm_ptr.is_null() {
+        return Err("ndk_context: JavaVM pointer is null".into());
+    }
+    let vm = unsafe { JavaVM::from_raw(vm_ptr) }
+        .map_err(|e| format!("JavaVM::from_raw: {e}"))?;
+    let activity_ptr = ctx.context() as jni::sys::jobject;
+    if activity_ptr.is_null() {
+        return Err("ndk_context: activity pointer is null".into());
+    }
+    // SAFETY: ndk_context guarantees the pointer lives for the process
+    // lifetime; we wrap it as a JObject<'static> for convenience.
+    let activity: JObject<'static> = unsafe { JObject::from_raw(activity_ptr) };
+    Ok((vm, activity))
+}
+
+/// Get Android's `AudioManager` via `activity.getSystemService("audio")`.
+fn audio_manager<'local>(
+    env: &mut jni::AttachGuard<'local>,
+    activity: &JObject<'local>,
+) -> Result<JObject<'local>, String> {
+    let svc_name: JString<'local> = env
+        .new_string("audio")
+        .map_err(|e| format!("new_string(audio): {e}"))?;
+    let am = env
+        .call_method(
+            activity,
+            "getSystemService",
+            "(Ljava/lang/String;)Ljava/lang/Object;",
+            &[JValue::Object(&svc_name)],
+        )
+        .and_then(|v| v.l())
+        .map_err(|e| format!("getSystemService(audio): {e}"))?;
+    if am.is_null() {
+        return Err("getSystemService returned null".into());
+    }
+    Ok(am)
+}
+
+/// Set `AudioManager.MODE_IN_COMMUNICATION`. Call when a VoIP call starts.
+/// This tells the audio policy to route through the communication device
+/// path (earpiece/BT SCO) instead of the media path (speaker/BT A2DP).
+pub fn set_audio_mode_communication() -> Result<(), String> {
+    let (vm, activity) = jvm_and_activity()?;
+    let mut env = vm
+        .attach_current_thread()
+        .map_err(|e| format!("attach_current_thread: {e}"))?;
+    let am = audio_manager(&mut env, &activity)?;
+    // MODE_IN_COMMUNICATION = 3
+    env.call_method(&am, "setMode", "(I)V", &[JValue::Int(3)])
+        .map_err(|e| format!("setMode(MODE_IN_COMMUNICATION): {e}"))?;
+    tracing::info!("AudioManager: mode set to MODE_IN_COMMUNICATION");
+    Ok(())
+}
+
+/// Restore `AudioManager.MODE_NORMAL`. Call when a VoIP call ends.
+pub fn set_audio_mode_normal() -> Result<(), String> {
+    let (vm, activity) = jvm_and_activity()?;
+    let mut env = vm
+        .attach_current_thread()
+        .map_err(|e| format!("attach_current_thread: {e}"))?;
+    let am = audio_manager(&mut env, &activity)?;
+    // MODE_NORMAL = 0
+    env.call_method(&am, "setMode", "(I)V", &[JValue::Int(0)])
+        .map_err(|e| format!("setMode(MODE_NORMAL): {e}"))?;
+    tracing::info!("AudioManager: mode set to MODE_NORMAL");
+    Ok(())
+}
+
+/// Switch between loud speaker (`true`) and earpiece/handset (`false`).
+pub fn set_speakerphone(on: bool) -> Result<(), String> {
+    let (vm, activity) = jvm_and_activity()?;
+    let mut env = vm
+        .attach_current_thread()
+        .map_err(|e| format!("attach_current_thread: {e}"))?;
+    let am = audio_manager(&mut env, &activity)?;
+
+    env.call_method(
+        &am,
+        "setSpeakerphoneOn",
+        "(Z)V",
+        &[JValue::Bool(if on { 1 } else { 0 })],
+    )
+    .map_err(|e| format!("setSpeakerphoneOn({on}): {e}"))?;
+
+    tracing::info!(on, "AudioManager.setSpeakerphoneOn");
+    Ok(())
+}
+
+/// Query the current speakerphone state. Returns true if routing is on the
+/// loud speaker, false if on earpiece / BT headset / wired headset.
+pub fn is_speakerphone_on() -> Result<bool, String> {
+    let (vm, activity) = jvm_and_activity()?;
+    let mut env = vm
+        .attach_current_thread()
+        .map_err(|e| format!("attach_current_thread: {e}"))?;
+    let am = audio_manager(&mut env, &activity)?;
+
+    let on = env
+        .call_method(&am, "isSpeakerphoneOn", "()Z", &[])
+        .and_then(|v| v.z())
+        .map_err(|e| format!("isSpeakerphoneOn: {e}"))?;
+    Ok(on)
+}
+
+// ─── Bluetooth SCO routing ──────────────────────────────────────────────────
+
+/// Start Bluetooth SCO audio routing.
+///
+/// On API 31+ uses `setCommunicationDevice()` which is the modern way to
+/// route voice audio to a specific device. Falls back to the deprecated
+/// `startBluetoothSco()` path on older APIs.
+///
+/// The caller must restart Oboe streams after this call.
+pub fn start_bluetooth_sco() -> Result<(), String> {
+    let (vm, activity) = jvm_and_activity()?;
+    let mut env = vm
+        .attach_current_thread()
+        .map_err(|e| format!("attach_current_thread: {e}"))?;
+    let am = audio_manager(&mut env, &activity)?;
+
+    // Ensure speaker is off — mutually exclusive with BT.
+    env.call_method(
+        &am,
+        "setSpeakerphoneOn",
+        "(Z)V",
+        &[JValue::Bool(0)],
+    )
+    .map_err(|e| format!("setSpeakerphoneOn(false): {e}"))?;
+
+    // Try modern API first (API 31+): setCommunicationDevice(AudioDeviceInfo)
+    // Find a BT SCO or BLE device from getAvailableCommunicationDevices()
+    let used_modern = try_set_communication_device(&mut env, &am, true)?;
+
+    if !used_modern {
+        // Fallback: deprecated startBluetoothSco (API < 31)
+        tracing::info!("start_bluetooth_sco: falling back to deprecated startBluetoothSco");
+        env.call_method(&am, "startBluetoothSco", "()V", &[])
+            .map_err(|e| format!("startBluetoothSco: {e}"))?;
+    }
+
+    tracing::info!(used_modern, "AudioManager: Bluetooth SCO started");
+    Ok(())
+}
+
+/// Stop Bluetooth SCO audio routing, returning audio to the earpiece.
+///
+/// The caller must restart Oboe streams after this call.
+pub fn stop_bluetooth_sco() -> Result<(), String> {
+    let (vm, activity) = jvm_and_activity()?;
+    let mut env = vm
+        .attach_current_thread()
+        .map_err(|e| format!("attach_current_thread: {e}"))?;
+    let am = audio_manager(&mut env, &activity)?;
+
+    // Modern API: clearCommunicationDevice() (API 31+)
+    let cleared = try_set_communication_device(&mut env, &am, false)?;
+
+    if !cleared {
+        // Fallback: deprecated stopBluetoothSco
+        env.call_method(&am, "stopBluetoothSco", "()V", &[])
+            .map_err(|e| format!("stopBluetoothSco: {e}"))?;
+    }
+
+    tracing::info!(cleared, "AudioManager: Bluetooth SCO stopped");
+    Ok(())
+}
+
+/// Try to use the modern `setCommunicationDevice` / `clearCommunicationDevice`
+/// API (Android 12 / API 31+). Returns `true` if the modern API was used.
+fn try_set_communication_device(
+    env: &mut jni::AttachGuard<'_>,
+    am: &JObject<'_>,
+    enable: bool,
+) -> Result<bool, String> {
+    // Check SDK_INT >= 31 (Android 12)
+    let sdk_int = env
+        .get_static_field(
+            "android/os/Build$VERSION",
+            "SDK_INT",
+            "I",
+        )
+        .and_then(|v| v.i())
+        .unwrap_or(0);
+
+    if sdk_int < 31 {
+        return Ok(false);
+    }
+
+    if !enable {
+        // clearCommunicationDevice()
+        env.call_method(am, "clearCommunicationDevice", "()V", &[])
+            .map_err(|e| format!("clearCommunicationDevice: {e}"))?;
+        tracing::info!("clearCommunicationDevice: done");
+        return Ok(true);
+    }
+
+    // getAvailableCommunicationDevices() → List<AudioDeviceInfo>
+    let device_list = env
+        .call_method(
+            am,
+            "getAvailableCommunicationDevices",
+            "()Ljava/util/List;",
+            &[],
+        )
+        .and_then(|v| v.l())
+        .map_err(|e| format!("getAvailableCommunicationDevices: {e}"))?;
+
+    let size = env
+        .call_method(&device_list, "size", "()I", &[])
+        .and_then(|v| v.i())
+        .unwrap_or(0);
+
+    // Find first BT device: TYPE_BLUETOOTH_SCO (7), TYPE_BLUETOOTH_A2DP (8),
+    // TYPE_BLE_HEADSET (26), TYPE_BLE_SPEAKER (27)
+    for i in 0..size {
+        let device = env
+            .call_method(
+                &device_list,
+                "get",
+                "(I)Ljava/lang/Object;",
+                &[JValue::Int(i)],
+            )
+            .and_then(|v| v.l())
+            .map_err(|e| format!("list.get({i}): {e}"))?;
+
+        let device_type = env
+            .call_method(&device, "getType", "()I", &[])
+            .and_then(|v| v.i())
+            .unwrap_or(0);
+
+        // BT SCO = 7, A2DP = 8, BLE headset = 26, BLE speaker = 27
+        if matches!(device_type, 7 | 8 | 26 | 27) {
+            let ok = env
+                .call_method(
+                    am,
+                    "setCommunicationDevice",
+                    "(Landroid/media/AudioDeviceInfo;)Z",
+                    &[JValue::Object(&device)],
+                )
+                .and_then(|v| v.z())
+                .unwrap_or(false);
+
+            tracing::info!(
+                device_type,
+                ok,
+                "setCommunicationDevice: set BT device"
+            );
+            return Ok(ok);
+        }
+    }
+
+    tracing::warn!("setCommunicationDevice: no BT device in available list");
+    Ok(false)
+}
+
+/// Query whether Bluetooth audio is currently the active communication device.
+///
+/// On API 31+ checks `getCommunicationDevice()` type. Falls back to the
+/// deprecated `isBluetoothScoOn()` on older APIs.
+pub fn is_bluetooth_sco_on() -> Result<bool, String> {
+    let (vm, activity) = jvm_and_activity()?;
+    let mut env = vm
+        .attach_current_thread()
+        .map_err(|e| format!("attach_current_thread: {e}"))?;
+    let am = audio_manager(&mut env, &activity)?;
+
+    let sdk_int = env
+        .get_static_field("android/os/Build$VERSION", "SDK_INT", "I")
+        .and_then(|v| v.i())
+        .unwrap_or(0);
+
+    if sdk_int >= 31 {
+        // getCommunicationDevice() → AudioDeviceInfo (nullable)
+        let device = env
+            .call_method(am, "getCommunicationDevice", "()Landroid/media/AudioDeviceInfo;", &[])
+            .and_then(|v| v.l())
+            .unwrap_or(JObject::null());
+        if device.is_null() {
+            return Ok(false);
+        }
+        let device_type = env
+            .call_method(&device, "getType", "()I", &[])
+            .and_then(|v| v.i())
+            .unwrap_or(0);
+        // BT SCO = 7, A2DP = 8, BLE headset = 26, BLE speaker = 27
+        return Ok(matches!(device_type, 7 | 8 | 26 | 27));
+    }
+
+    // Fallback: deprecated API
+    env.call_method(&am, "isBluetoothScoOn", "()Z", &[])
+        .and_then(|v| v.z())
+        .map_err(|e| format!("isBluetoothScoOn: {e}"))
+}
+
+/// Check whether a Bluetooth audio device is currently connected.
+///
+/// Iterates `AudioManager.getDevices(GET_DEVICES_OUTPUTS)` and looks for
+/// any Bluetooth device type. Many headsets only register as A2DP until
+/// SCO is explicitly started, so we check for both SCO and A2DP types.
+pub fn is_bluetooth_available() -> Result<bool, String> {
+    let (vm, activity) = jvm_and_activity()?;
+    let mut env = vm
+        .attach_current_thread()
+        .map_err(|e| format!("attach_current_thread: {e}"))?;
+    let am = audio_manager(&mut env, &activity)?;
+
+    // AudioManager.GET_DEVICES_OUTPUTS = 2
+    let devices = env
+        .call_method(
+            &am,
+            "getDevices",
+            "(I)[Landroid/media/AudioDeviceInfo;",
+            &[JValue::Int(2)],
+        )
+        .and_then(|v| v.l())
+        .map_err(|e| format!("getDevices(OUTPUTS): {e}"))?;
+
+    let arr = jni::objects::JObjectArray::from(devices);
+    let len = env
+        .get_array_length(&arr)
+        .map_err(|e| format!("get_array_length: {e}"))?;
+
+    for i in 0..len {
+        let device = env
+            .get_object_array_element(&arr, i)
+            .map_err(|e| format!("get_object_array_element({i}): {e}"))?;
+        let device_type = env
+            .call_method(&device, "getType", "()I", &[])
+            .and_then(|v| v.i())
+            .unwrap_or(0);
+        // TYPE_BLUETOOTH_SCO = 7, TYPE_BLUETOOTH_A2DP = 8
+        if device_type == 7 || device_type == 8 {
+            tracing::info!(device_type, idx = i, "is_bluetooth_available: found BT device");
+            return Ok(true);
+        }
+    }
+    Ok(false)
+}