test+docs: add tests for QualityDirective, ParticipantQuality; update docs

- QualityDirective signal roundtrip tests (with/without reason)
- ParticipantQuality unit tests (initial tier, degradation, weakest-link)
- Updated PROGRESS.md with desktop adaptive quality, relay coordinated
  switching, Oboe state polling entries
- Updated ARCHITECTURE.md SFU fan-out rules with QualityDirective
- Updated PRD-coordinated-codec.md with implementation status
- 312 tests passing across all modified crates

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,
                 },
@@ -709,14 +738,16 @@ 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 !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) => {
@@ -767,6 +798,7 @@ async fn run_call(
                     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
+                    // 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]);
@@ -899,12 +1085,21 @@ 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
+                    // 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 {
@@ -921,10 +1116,13 @@ async fn run_call(
                         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(),

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

@@ -23,10 +23,77 @@ serde_json = "1"
 chrono = "0.4"
 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"
+
+# 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"

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/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,
         }
     }
 }
@@ -241,7 +246,7 @@ 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,
@@ -340,6 +345,22 @@ 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 {
@@ -347,11 +368,11 @@ impl CallEncoder {
                 is_repair: false,
                 codec_id: self.profile.codec,
                 has_quality_report: false,
-                fec_ratio_encoded: MediaHeader::encode_fec_ratio(self.profile.fec_ratio),
+                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,
             },
@@ -366,11 +387,13 @@ impl CallEncoder {
 
         let mut output = vec![source_pkt];
 
-        // Add to FEC encoder
+        // 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 {
@@ -400,6 +423,7 @@ impl CallEncoder {
                 self.block_id = self.block_id.wrapping_add(1);
                 self.frame_in_block = 0;
             }
+        }
 
         Ok(output)
     }
@@ -421,6 +445,15 @@ 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);
+    }
+
     /// Enable or disable acoustic echo cancellation.
     pub fn set_aec_enabled(&mut self, enabled: bool) {
         self.aec.set_enabled(enabled);
@@ -434,9 +467,12 @@ 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 +486,24 @@ 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 +513,19 @@ 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 +533,12 @@ 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 +553,105 @@ impl CallDecoder {
 
     /// Feed a received media packet into the decode pipeline.
     pub fn ingest(&mut self, packet: MediaPacket) {
-        // Feed to FEC decoder
+        // 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 +666,9 @@ 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 +683,72 @@ 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");
+                if self.jitter.depth() == 0 {
+                    self.jitter.record_underrun();
+                    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
-                } else {
-                    self.jitter.record_underrun();
-                    None
-                }
             }
             PlayoutResult::NotReady => {
                 self.jitter.record_underrun();
@@ -559,6 +771,19 @@ 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 +845,83 @@ 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 +930,10 @@ 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 +964,219 @@ 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 +1451,131 @@ 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());
+    }
 }

crates/wzp-client/src/cli.rs

@@ -424,6 +424,7 @@ 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 +576,7 @@ 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 +628,21 @@ 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 +673,13 @@ 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 +749,7 @@ 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 +772,17 @@ 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_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 +806,17 @@ 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_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: _ } => {
                     info!(call_id = %call_id, room = %room, relay = %setup_relay, "call setup — connecting to media room");
 
                     // Connect to the media room
@@ -840,6 +867,7 @@ 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 +884,7 @@ 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 +1,546 @@
+//! 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,
+}
+
+/// 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,
+}
+
+/// 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>,
+}
+
+impl PeerCandidates {
+    /// Flatten into the list of addrs the D-role should dial.
+    /// Order: LAN host candidates first (fastest when they
+    /// work), then reflexive (covers the non-LAN case).
+    pub fn dial_order(&self) -> Vec<SocketAddr> {
+        let mut out = Vec::with_capacity(self.local.len() + 1);
+        out.extend(self.local.iter().copied());
+        if let Some(a) = self.reflexive {
+            // Only add if it's not already in the list (some
+            // edge cases on same-LAN could have the same addr
+            // in both).
+            if !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()
+    }
+}
+
+#[allow(clippy::too_many_arguments)]
+pub async fn race(
+    role: Role,
+    peer_candidates: PeerCandidates,
+    relay_addr: SocketAddr,
+    room_sni: String,
+    call_sni: String,
+    // 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();
+
+    // 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();
+            direct_fut = Box::pin(async move {
+                // 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.dial_order();
+            let sni = call_sni.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::debug!(
+                                %candidate,
+                                candidate_idx = idx,
+                                "dual_path: skipping IPv6 candidate (disabled)"
+                            );
+                            continue;
+                        }
+                        let ep = ep_for_fut.clone();
+                        let client_cfg = wzp_transport::client_config();
+                        let sni = sni.clone();
+                        set.spawn(async move {
+                            let result = wzp_transport::connect(
+                                &ep,
+                                candidate,
+                                &sni,
+                                client_cfg,
+                            )
+                            .await;
+                            (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::debug!(
+                                    %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.
+    tracing::info!(
+        ?role,
+        candidates = ?peer_candidates.dial_order(),
+        %relay_addr,
+        "dual_path: racing direct vs relay"
+    );
+
+    let mut direct_task = tokio::spawn(
+        tokio::time::timeout(Duration::from_secs(2), 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 (2s)");
+                    direct_result = Some(Err(anyhow::anyhow!("direct timeout")));
+                    local_winner = WinningPath::Relay;
+                }
+                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"))); }
+        }
+    }
+    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);
+
+    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,
+    })
+}

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 +120,18 @@ 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::QualityDirective { .. } => CallSignalType::Offer, // relay-initiated
     }
 }
 
@@ -158,6 +171,7 @@ 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/lib.rs

@@ -8,16 +8,77 @@
 
 #[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 reflect;
 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/reflect.rs

@@ -0,0 +1,679 @@
+//! 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)
+    }
+}
+
+// ── 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/tests/dual_path.rs

@@ -0,0 +1,213 @@
+//! 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(),
+        },
+        relay_addr,
+        "test-room".into(),
+        "call-test".into(),
+        None, // Phase 5: tests use fresh endpoints (no shared signal)
+    )
+    .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(),
+        },
+        relay_addr,
+        "test-room".into(),
+        "call-test".into(),
+        None, // Phase 5: tests use fresh endpoints (no shared signal)
+    )
+    .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(),
+        },
+        dead_relay,
+        "test-room".into(),
+        "call-test".into(),
+        None, // Phase 5: tests use fresh endpoints (no shared signal)
+    )
+    .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;
-
+            // --- 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;
-                let fe = self.far_end_buf[fe_idx];
-                self.filter_coeffs[k] += step * fe;
+                    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
         }
+    }
 
-    /// 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_bitrate(Bitrate::BitsPerSecond(bps))
-            .map_err(|e| CodecError::EncodeFailed(format!("set bitrate: {e}")))?;
+                .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_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/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 +1,427 @@
+//! 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.
+#[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;
+        }
+    }
+}
+
+/// 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).
+#[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).
+#[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 +1,312 @@
+//! 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.
+
+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;
@@ -30,6 +31,7 @@ pub use packet::{
     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/packet.rs

@@ -584,12 +584,38 @@ pub enum SignalMessage {
         recommended_profile: crate::QualityProfile,
     },
 
+    /// Phase 4 telemetry: loss-recovery counts for the current session.
+    /// Sent periodically from receivers to the relay so Prometheus metrics
+    /// can distinguish DRED reconstructions from classical PLC invocations.
+    /// Fields default to 0 on old receivers (`#[serde(default)]`), so
+    /// introducing this variant is backward-compatible with pre-Phase-4
+    /// relays — they'll just log "unknown signal variant" on receipt.
+    LossRecoveryUpdate {
+        /// Total frames reconstructed via DRED since call start (monotonic).
+        #[serde(default)]
+        dred_reconstructions: u64,
+        /// Total frames filled via classical Opus/Codec2 PLC since call
+        /// start (monotonic).
+        #[serde(default)]
+        classical_plc_invocations: u64,
+        /// Total frames decoded since call start. Used by the relay to
+        /// compute recovery rates as a fraction of total frames.
+        #[serde(default)]
+        frames_decoded: u64,
+    },
+
     /// Connection keepalive / RTT measurement.
     Ping { timestamp_ms: u64 },
     Pong { timestamp_ms: u64 },
 
-    /// End the call.
-    Hangup { reason: HangupReason },
+    /// End the call. `call_id` is optional for backwards compatibility
+    /// with older clients that send Hangup without it — the relay falls
+    /// back to ending ALL active calls for the sender in that case.
+    Hangup {
+        reason: HangupReason,
+        #[serde(default, skip_serializing_if = "Option::is_none")]
+        call_id: Option<String>,
+    },
 
     /// featherChat bearer token for relay authentication.
     /// Sent as the first signal message when --auth-url is configured.
@@ -696,6 +722,9 @@ pub enum SignalMessage {
         success: bool,
         #[serde(skip_serializing_if = "Option::is_none")]
         error: Option<String>,
+        /// Relay's build version (git short hash).
+        #[serde(default, skip_serializing_if = "Option::is_none")]
+        relay_build: Option<String>,
     },
 
     /// Direct call offer routed through the relay to a specific peer.
@@ -716,6 +745,28 @@ pub enum SignalMessage {
         signature: Vec<u8>,
         /// Supported quality profiles.
         supported_profiles: Vec<crate::QualityProfile>,
+        /// Phase 3 (hole-punching): caller's own server-reflexive
+        /// address as learned via `SignalMessage::Reflect`. The
+        /// relay stashes this in its call registry and later
+        /// injects it into the callee's `CallSetup.peer_direct_addr`
+        /// so the callee can try a direct QUIC handshake to the
+        /// caller instead of routing media through the relay.
+        /// `None` means "caller doesn't want P2P, use relay only".
+        #[serde(default, skip_serializing_if = "Option::is_none")]
+        caller_reflexive_addr: Option<String>,
+        /// Phase 5.5 (ICE host candidates): caller's LAN-local
+        /// interface addresses paired with its signal endpoint's
+        /// port. Peers on the same physical LAN can direct-dial
+        /// these without going through the WAN reflex addr,
+        /// which is important because most consumer NATs
+        /// (including MikroTik masquerade) don't support NAT
+        /// hairpinning — the reflex addr is unreachable from
+        /// the same LAN.
+        #[serde(default, skip_serializing_if = "Vec::is_empty")]
+        caller_local_addrs: Vec<String>,
+        /// Build version (git short hash) for debugging.
+        #[serde(default, skip_serializing_if = "Option::is_none")]
+        caller_build_version: Option<String>,
     },
 
     /// Callee's response to a direct call.
@@ -735,6 +786,23 @@ pub enum SignalMessage {
         /// Chosen quality profile (present when accepting).
         #[serde(skip_serializing_if = "Option::is_none")]
         chosen_profile: Option<crate::QualityProfile>,
+        /// Phase 3 (hole-punching): callee's own server-reflexive
+        /// address, only populated on `AcceptTrusted` — privacy-mode
+        /// answers leave this `None` so the callee's real IP stays
+        /// hidden (the whole point of `AcceptGeneric`). The relay
+        /// carries it opaquely into the caller's `CallSetup`.
+        #[serde(default, skip_serializing_if = "Option::is_none")]
+        callee_reflexive_addr: Option<String>,
+        /// Phase 5.5 (ICE host candidates): callee's LAN-local
+        /// interface addresses. Same purpose as
+        /// `caller_local_addrs` in `DirectCallOffer`. Only
+        /// populated on `AcceptTrusted` alongside
+        /// `callee_reflexive_addr`.
+        #[serde(default, skip_serializing_if = "Vec::is_empty")]
+        callee_local_addrs: Vec<String>,
+        /// Build version (git short hash) for debugging.
+        #[serde(default, skip_serializing_if = "Option::is_none")]
+        callee_build_version: Option<String>,
     },
 
     /// Relay tells both parties: media room is ready.
@@ -744,12 +812,119 @@ pub enum SignalMessage {
         room: String,
         /// Relay address for the QUIC media connection.
         relay_addr: String,
+        /// Phase 3 (hole-punching): the OTHER party's server-reflexive
+        /// address as the relay learned it from the offer/answer
+        /// exchange. When populated, clients attempt a direct QUIC
+        /// handshake to this address in parallel with the existing
+        /// relay path and use whichever connects first. `None`
+        /// means the relay path is the only option — either because
+        /// a peer didn't advertise its addr (Phase 1/2 relay or
+        /// privacy-mode answer) or because the relay decided P2P
+        /// wasn't viable.
+        #[serde(default, skip_serializing_if = "Option::is_none")]
+        peer_direct_addr: Option<String>,
+        /// Phase 5.5 (ICE host candidates): the OTHER party's LAN
+        /// host addresses (RFC1918 IPv4 + CGNAT + non-link-local
+        /// IPv6). On same-LAN calls these are directly dialable
+        /// and bypass the NAT-hairpinning problem that blocks
+        /// same-LAN peers from using `peer_direct_addr`.
+        /// Client-side race tries all of these in parallel.
+        #[serde(default, skip_serializing_if = "Vec::is_empty")]
+        peer_local_addrs: Vec<String>,
     },
 
     /// Ringing notification (relay → caller, callee received the offer).
     CallRinging {
         call_id: String,
     },
+
+    // ── NAT reflection ("STUN for QUIC") ──────────────────────────────
+
+    /// Client → relay: "please tell me the source IP:port you see on
+    /// this connection". A QUIC-native replacement for classic STUN
+    /// that reuses the TLS-authenticated signal channel to the relay
+    /// instead of running a separate UDP reflection service on port
+    /// 3478. The relay answers with `ReflectResponse`.
+    ///
+    /// No payload — the relay already knows which connection the
+    /// request arrived on, and `connection.remote_address()` gives it
+    /// the exact source address (post-NAT) as observed from the
+    /// server side of the TLS session.
+    Reflect,
+
+    /// Relay → client: response to `Reflect`. Carries the socket
+    /// address the relay observes as the client's source for this
+    /// QUIC connection in `SocketAddr::to_string()` form — "a.b.c.d:p"
+    /// for IPv4, "[::1]:p" for IPv6. Clients parse it with
+    /// `SocketAddr::from_str`.
+    ReflectResponse {
+        observed_addr: String,
+    },
+
+    // ── Phase 6: ICE-style path negotiation ─────────────────────
+
+    /// Phase 6: each side reports the result of its local dual-
+    /// path race to the other side through the relay. Both peers
+    /// send this after their race completes; both wait for the
+    /// other's report before committing a transport to the
+    /// CallEngine.
+    ///
+    /// The decision rule is: if BOTH sides report `direct_ok =
+    /// true`, use the direct P2P connection. If EITHER reports
+    /// `direct_ok = false`, BOTH fall back to relay. This
+    /// eliminates the race condition where one side picks Direct
+    /// and the other picks Relay — they now agree on the path
+    /// before any media flows.
+    MediaPathReport {
+        call_id: String,
+        /// Did the direct QUIC connection (P2P dial or accept)
+        /// complete successfully on this side?
+        direct_ok: bool,
+        /// Which future won the local tokio::select race?
+        /// "Direct" or "Relay" — informational for debug logs.
+        #[serde(default)]
+        race_winner: String,
+    },
+
+    // ── Phase 4: cross-relay direct-call signaling ────────────────────
+
+    /// Phase 4: relay-to-relay envelope for forwarding direct-call
+    /// signaling across a federation link. When Alice on Relay A
+    /// sends a `DirectCallOffer` for Bob whose fingerprint isn't
+    /// in A's local SignalHub, Relay A wraps the offer in this
+    /// envelope and broadcasts it over every active federation
+    /// peer link. Whichever peer has Bob registered unwraps the
+    /// inner message and delivers it locally.
+    ///
+    /// Never originated by clients — only relays create and
+    /// consume this variant.
+    ///
+    /// Loop prevention: the receiving relay drops any forward
+    /// where `origin_relay_fp` matches its own federation TLS
+    /// fingerprint. With broadcast-to-all-peers this prevents
+    /// A→B→A echo loops; proper TTL + dedup will land when
+    /// multi-hop federation is added (Phase 4.2).
+    FederatedSignalForward {
+        /// The signal message being forwarded
+        /// (`DirectCallOffer`, `DirectCallAnswer`, `CallRinging`,
+        /// `Hangup`, ...). Boxed because `SignalMessage` is
+        /// relatively large and JSON serde handles recursion
+        /// cleanly.
+        inner: Box<SignalMessage>,
+        /// Federation TLS fingerprint of the sending relay.
+        /// Used (a) for loop prevention by the receiver and (b)
+        /// to route the peer's reply back through the same
+        /// federation link via `send_signal_to_peer`.
+        origin_relay_fp: String,
+    },
+
+    /// Relay-initiated quality directive: all participants should switch
+    /// to the recommended profile to match the weakest link.
+    QualityDirective {
+        recommended_profile: crate::QualityProfile,
+        #[serde(default, skip_serializing_if = "Option::is_none")]
+        reason: Option<String>,
+    },
 }
 
 /// How the callee responds to a direct call.
@@ -888,6 +1063,272 @@ mod tests {
         assert_eq!(packet.quality_report, decoded.quality_report);
     }
 
+    #[test]
+    fn reflect_serialize_roundtrip() {
+        // Reflect is a unit variant — the client sends it with no
+        // payload and the relay answers with the observed source addr.
+        let req = SignalMessage::Reflect;
+        let json = serde_json::to_string(&req).unwrap();
+        let decoded: SignalMessage = serde_json::from_str(&json).unwrap();
+        assert!(matches!(decoded, SignalMessage::Reflect));
+
+        // ReflectResponse carries a string — exercise both IPv4 and
+        // IPv6 shapes because SocketAddr::to_string uses [::1]:port
+        // for v6 and the client side has to parse that back.
+        for addr in ["192.0.2.17:4433", "[2001:db8::1]:4433", "127.0.0.1:54321"] {
+            let resp = SignalMessage::ReflectResponse {
+                observed_addr: addr.to_string(),
+            };
+            let json = serde_json::to_string(&resp).unwrap();
+            let decoded: SignalMessage = serde_json::from_str(&json).unwrap();
+            match decoded {
+                SignalMessage::ReflectResponse { observed_addr } => {
+                    assert_eq!(observed_addr, addr);
+                    // Must parse back to a SocketAddr cleanly.
+                    let _parsed: std::net::SocketAddr = observed_addr.parse()
+                        .expect("observed_addr must parse as SocketAddr");
+                }
+                _ => panic!("wrong variant after roundtrip"),
+            }
+        }
+    }
+
+    #[test]
+    fn federated_signal_forward_roundtrip() {
+        // Wrap a DirectCallOffer inside FederatedSignalForward and
+        // prove both directions of serde preserve every field.
+        let inner = SignalMessage::DirectCallOffer {
+            caller_fingerprint: "alice".into(),
+            caller_alias: Some("Alice".into()),
+            target_fingerprint: "bob".into(),
+            call_id: "c1".into(),
+            identity_pub: [1u8; 32],
+            ephemeral_pub: [2u8; 32],
+            signature: vec![3u8; 64],
+            supported_profiles: vec![],
+            caller_reflexive_addr: Some("192.0.2.1:4433".into()),
+            caller_local_addrs: Vec::new(),
+            caller_build_version: None,
+        };
+        let forward = SignalMessage::FederatedSignalForward {
+            inner: Box::new(inner),
+            origin_relay_fp: "relay-a-tls-fp".into(),
+        };
+        let json = serde_json::to_string(&forward).unwrap();
+        let decoded: SignalMessage = serde_json::from_str(&json).unwrap();
+        match decoded {
+            SignalMessage::FederatedSignalForward { inner, origin_relay_fp } => {
+                assert_eq!(origin_relay_fp, "relay-a-tls-fp");
+                match *inner {
+                    SignalMessage::DirectCallOffer {
+                        caller_fingerprint,
+                        target_fingerprint,
+                        caller_reflexive_addr,
+                        ..
+                    } => {
+                        assert_eq!(caller_fingerprint, "alice");
+                        assert_eq!(target_fingerprint, "bob");
+                        assert_eq!(caller_reflexive_addr.as_deref(), Some("192.0.2.1:4433"));
+                    }
+                    _ => panic!("inner was not DirectCallOffer after roundtrip"),
+                }
+            }
+            _ => panic!("outer was not FederatedSignalForward"),
+        }
+    }
+
+    #[test]
+    fn federated_signal_forward_can_nest_any_inner() {
+        // Sanity check that every direct-call signaling variant
+        // we intend to forward survives being boxed + re-serialized.
+        let cases: Vec<SignalMessage> = vec![
+            SignalMessage::DirectCallAnswer {
+                call_id: "c1".into(),
+                accept_mode: CallAcceptMode::AcceptTrusted,
+                identity_pub: None,
+                ephemeral_pub: None,
+                signature: None,
+                chosen_profile: None,
+                callee_reflexive_addr: Some("198.51.100.9:4433".into()),
+                callee_local_addrs: Vec::new(),
+                callee_build_version: None,
+            },
+            SignalMessage::CallRinging { call_id: "c1".into() },
+            SignalMessage::Hangup { reason: HangupReason::Normal, call_id: None },
+        ];
+        for inner in cases {
+            let inner_disc = std::mem::discriminant(&inner);
+            let forward = SignalMessage::FederatedSignalForward {
+                inner: Box::new(inner),
+                origin_relay_fp: "r".into(),
+            };
+            let json = serde_json::to_string(&forward).unwrap();
+            let decoded: SignalMessage = serde_json::from_str(&json).unwrap();
+            match decoded {
+                SignalMessage::FederatedSignalForward { inner, .. } => {
+                    assert_eq!(std::mem::discriminant(&*inner), inner_disc);
+                }
+                _ => panic!("outer variant lost"),
+            }
+        }
+    }
+
+    #[test]
+    fn hole_punching_optional_fields_roundtrip() {
+        // DirectCallOffer with Some(caller_reflexive_addr)
+        let offer = SignalMessage::DirectCallOffer {
+            caller_fingerprint: "alice".into(),
+            caller_alias: None,
+            target_fingerprint: "bob".into(),
+            call_id: "c1".into(),
+            identity_pub: [0; 32],
+            ephemeral_pub: [0; 32],
+            signature: vec![],
+            supported_profiles: vec![],
+            caller_reflexive_addr: Some("192.0.2.1:4433".into()),
+            caller_local_addrs: Vec::new(),
+            caller_build_version: None,
+        };
+        let json = serde_json::to_string(&offer).unwrap();
+        assert!(
+            json.contains("caller_reflexive_addr"),
+            "Some field must serialize: {json}"
+        );
+        let decoded: SignalMessage = serde_json::from_str(&json).unwrap();
+        match decoded {
+            SignalMessage::DirectCallOffer { caller_reflexive_addr, .. } => {
+                assert_eq!(caller_reflexive_addr.as_deref(), Some("192.0.2.1:4433"));
+            }
+            _ => panic!("wrong variant"),
+        }
+
+        // DirectCallOffer with None — skip_serializing_if must
+        // OMIT the field from the JSON so older relays that don't
+        // know about caller_reflexive_addr don't see it.
+        let offer_none = SignalMessage::DirectCallOffer {
+            caller_fingerprint: "alice".into(),
+            caller_alias: None,
+            target_fingerprint: "bob".into(),
+            call_id: "c1".into(),
+            identity_pub: [0; 32],
+            ephemeral_pub: [0; 32],
+            signature: vec![],
+            supported_profiles: vec![],
+            caller_reflexive_addr: None,
+            caller_local_addrs: Vec::new(),
+            caller_build_version: None,
+        };
+        let json_none = serde_json::to_string(&offer_none).unwrap();
+        assert!(
+            !json_none.contains("caller_reflexive_addr"),
+            "None field must NOT serialize: {json_none}"
+        );
+
+        // DirectCallAnswer with callee_reflexive_addr.
+        let answer = SignalMessage::DirectCallAnswer {
+            call_id: "c1".into(),
+            accept_mode: CallAcceptMode::AcceptTrusted,
+            identity_pub: None,
+            ephemeral_pub: None,
+            signature: None,
+            chosen_profile: None,
+            callee_reflexive_addr: Some("198.51.100.9:4433".into()),
+            callee_local_addrs: Vec::new(),
+            callee_build_version: None,
+        };
+        let decoded: SignalMessage =
+            serde_json::from_str(&serde_json::to_string(&answer).unwrap()).unwrap();
+        match decoded {
+            SignalMessage::DirectCallAnswer { callee_reflexive_addr, .. } => {
+                assert_eq!(
+                    callee_reflexive_addr.as_deref(),
+                    Some("198.51.100.9:4433")
+                );
+            }
+            _ => panic!("wrong variant"),
+        }
+
+        // CallSetup with peer_direct_addr.
+        let setup = SignalMessage::CallSetup {
+            call_id: "c1".into(),
+            room: "call-c1".into(),
+            relay_addr: "203.0.113.5:4433".into(),
+            peer_direct_addr: Some("192.0.2.1:4433".into()),
+            peer_local_addrs: Vec::new(),
+        };
+        let decoded: SignalMessage =
+            serde_json::from_str(&serde_json::to_string(&setup).unwrap()).unwrap();
+        match decoded {
+            SignalMessage::CallSetup { peer_direct_addr, .. } => {
+                assert_eq!(peer_direct_addr.as_deref(), Some("192.0.2.1:4433"));
+            }
+            _ => panic!("wrong variant"),
+        }
+    }
+
+    #[test]
+    fn hole_punching_backward_compat_old_json_parses() {
+        // An older client/relay wouldn't include the new fields at
+        // all — the new code must still accept that JSON because
+        // of #[serde(default)] on the Option<String>.
+        let old_offer_json = r#"{
+            "DirectCallOffer": {
+                "caller_fingerprint": "alice",
+                "caller_alias": null,
+                "target_fingerprint": "bob",
+                "call_id": "c1",
+                "identity_pub": [0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],
+                "ephemeral_pub": [0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],
+                "signature": [],
+                "supported_profiles": []
+            }
+        }"#;
+        let decoded: SignalMessage = serde_json::from_str(old_offer_json).unwrap();
+        match decoded {
+            SignalMessage::DirectCallOffer { caller_reflexive_addr, .. } => {
+                assert!(caller_reflexive_addr.is_none());
+            }
+            _ => panic!("wrong variant"),
+        }
+
+        let old_setup_json = r#"{
+            "CallSetup": {
+                "call_id": "c1",
+                "room": "call-c1",
+                "relay_addr": "203.0.113.5:4433"
+            }
+        }"#;
+        let decoded: SignalMessage = serde_json::from_str(old_setup_json).unwrap();
+        match decoded {
+            SignalMessage::CallSetup { peer_direct_addr, .. } => {
+                assert!(peer_direct_addr.is_none());
+            }
+            _ => panic!("wrong variant"),
+        }
+    }
+
+    #[test]
+    fn reflect_backward_compat_with_existing_variants() {
+        // Adding Reflect/ReflectResponse at the end of the enum must
+        // not break JSON round-tripping of existing variants. Smoke-
+        // test a sample of the pre-existing ones.
+        let cases = vec![
+            SignalMessage::Ping { timestamp_ms: 12345 },
+            SignalMessage::Hold,
+            SignalMessage::Hangup { reason: HangupReason::Normal, call_id: None },
+            SignalMessage::CallRinging { call_id: "abcd".into() },
+        ];
+        for m in cases {
+            let json = serde_json::to_string(&m).unwrap();
+            let decoded: SignalMessage = serde_json::from_str(&json).unwrap();
+            // Discriminant equality proves variant tag survived.
+            assert_eq!(
+                std::mem::discriminant(&m),
+                std::mem::discriminant(&decoded)
+            );
+        }
+    }
+
     #[test]
     fn hold_unhold_serialize() {
         let hold = SignalMessage::Hold;
@@ -1232,6 +1673,41 @@ mod tests {
         }
     }
 
+    #[test]
+    fn quality_directive_roundtrip() {
+        let msg = SignalMessage::QualityDirective {
+            recommended_profile: crate::QualityProfile::DEGRADED,
+            reason: Some("weakest link degraded".into()),
+        };
+        let json = serde_json::to_string(&msg).unwrap();
+        let decoded: SignalMessage = serde_json::from_str(&json).unwrap();
+        match decoded {
+            SignalMessage::QualityDirective { recommended_profile, reason } => {
+                assert_eq!(recommended_profile.codec, CodecId::Opus6k);
+                assert_eq!(reason.as_deref(), Some("weakest link degraded"));
+            }
+            _ => panic!("wrong variant"),
+        }
+    }
+
+    #[test]
+    fn quality_directive_without_reason_roundtrip() {
+        let msg = SignalMessage::QualityDirective {
+            recommended_profile: crate::QualityProfile::GOOD,
+            reason: None,
+        };
+        let json = serde_json::to_string(&msg).unwrap();
+        // None reason should be omitted from JSON
+        assert!(!json.contains("reason"));
+        let decoded: SignalMessage = serde_json::from_str(&json).unwrap();
+        match decoded {
+            SignalMessage::QualityDirective { reason, .. } => {
+                assert!(reason.is_none());
+            }
+            _ => panic!("wrong variant"),
+        }
+    }
+
     #[test]
     fn mini_frame_disabled() {
         // Simulate disabled mini-frames by always keeping frames_since_full at 0

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/src/call_registry.rs

@@ -31,6 +31,36 @@ 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>,
 }
 
 /// Registry of active direct calls.
@@ -57,11 +87,61 @@ 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(),
         };
         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;
+        }
+    }
+
     /// Get a call by ID.
     pub fn get(&self, call_id: &str) -> Option<&DirectCall> {
         self.calls.get(call_id)
@@ -196,4 +276,79 @@ 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_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/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

@@ -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 {
@@ -172,34 +177,133 @@ 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 links = self.peer_links.lock().await;
+        let mut count = 0;
+        for (fp, link) in links.iter() {
+            match link.transport.send_signal(msg).await {
+                Ok(()) => {
+                    count += 1;
+                    tracing::debug!(peer = %link.label, %fp, "federation: broadcast signal ok");
+                }
+                Err(e) => {
+                    tracing::warn!(peer = %link.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 links = self.peer_links.lock().await;
+        match links.get(&normalized) {
+            Some(link) => link
+                .transport
+                .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)
@@ -271,8 +375,8 @@ 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
@@ -296,7 +400,12 @@ 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) {
+    ///
+    /// `_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 links = self.peer_links.lock().await;
         if links.is_empty() {
             return;
@@ -623,11 +732,20 @@ 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);
         }
     };
 
@@ -717,12 +835,12 @@ async fn handle_signal(
                         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) {
+                            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 canonical != &local_room {
                                     if let Some(remote) = link.remote_participants.get(&local_room) {
                                         all_participants.extend(remote.iter().cloned());
                                     }
@@ -753,8 +871,8 @@ 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 +886,8 @@ 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());
                         }
                     }
@@ -842,6 +960,57 @@ async fn handle_signal(
                 }
             }
         }
+        // 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
     }
 }
@@ -908,7 +1077,7 @@ async fn handle_datagram(
         // 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 +1087,23 @@ 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 = {
+                let mgr = fm.room_mgr.lock().await;
+                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;
         }
     };

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/main.rs

@@ -13,7 +13,7 @@ use std::sync::Arc;
 use std::time::Duration;
 
 use tokio::sync::Mutex;
-use tracing::{error, info, warn};
+use tracing::{debug, error, info, warn};
 
 use wzp_proto::{MediaTransport, SignalMessage};
 use wzp_relay::config::RelayConfig;
@@ -272,7 +272,7 @@ const BUILD_GIT_HASH: &str = env!("WZP_BUILD_HASH");
 
 #[tokio::main]
 async fn main() -> anyhow::Result<()> {
-    let CliResult { mut config, identity_path, config_file, config_needs_create } = parse_args();
+    let CliResult { config, identity_path, config_file, config_needs_create } = parse_args();
     tracing_subscriber::fmt().init();
     info!(version = BUILD_GIT_HASH, "wzp-relay build");
     rustls::crypto::ring::default_provider()
@@ -378,6 +378,31 @@ async fn main() -> anyhow::Result<()> {
     }
     let endpoint = wzp_transport::create_endpoint(config.listen_addr, Some(server_config))?;
 
+    // Compute the IP address we should advertise in CallSetup for direct
+    // calls. If the relay is bound to a specific IP, use it as-is; if bound
+    // to 0.0.0.0, use the trick of "connect" a UDP socket to an arbitrary
+    // external address and read its local_addr — the OS binds to whichever
+    // local interface IP would route packets to that destination, which is
+    // the primary outbound interface. This is the same IP clients on the
+    // LAN use to reach us.
+    let advertised_ip: std::net::IpAddr = {
+        let listen_ip = config.listen_addr.ip();
+        if !listen_ip.is_unspecified() {
+            listen_ip
+        } else {
+            // Probe via a dummy "connected" UDP socket. Never actually sends.
+            match std::net::UdpSocket::bind("0.0.0.0:0")
+                .and_then(|s| { s.connect("8.8.8.8:80").map(|_| s) })
+                .and_then(|s| s.local_addr())
+            {
+                Ok(a) if !a.ip().is_loopback() => a.ip(),
+                _ => std::net::IpAddr::from([127u8, 0, 0, 1]),
+            }
+        }
+    };
+    let advertised_addr_str = format!("{}:{}", advertised_ip, config.listen_addr.port());
+    info!(%advertised_addr_str, "relay advertised address for CallSetup");
+
     // Forward mode
     let remote_transport: Option<Arc<wzp_transport::QuinnTransport>> =
         if let Some(remote_addr) = config.remote_relay {
@@ -428,6 +453,21 @@ async fn main() -> anyhow::Result<()> {
     let signal_hub = Arc::new(Mutex::new(wzp_relay::signal_hub::SignalHub::new()));
     let call_registry = Arc::new(Mutex::new(wzp_relay::call_registry::CallRegistry::new()));
 
+    // Phase 4: cross-relay direct-call signal dispatcher.
+    //
+    // The federation layer unwraps incoming
+    // `SignalMessage::FederatedSignalForward` envelopes and pushes
+    // (inner, origin_relay_fp) onto this channel. A dedicated task
+    // further down reads from it and routes the inner message
+    // through signal_hub / call_registry exactly as if it had
+    // arrived on a local signal transport — with the extra
+    // context that a peer relay is on the other side of the call.
+    let (cross_relay_tx, mut cross_relay_rx) =
+        tokio::sync::mpsc::channel::<(wzp_proto::SignalMessage, String)>(32);
+    if let Some(ref fm) = federation_mgr {
+        fm.set_cross_relay_tx(cross_relay_tx.clone()).await;
+    }
+
     // Spawn inter-relay health probes via ProbeMesh coordinator
     if !config.probe_targets.is_empty() {
         let mesh = wzp_relay::probe::ProbeMesh::new(
@@ -472,12 +512,256 @@ async fn main() -> anyhow::Result<()> {
         info!(filter = %tap, "debug tap enabled — logging packet headers");
     }
 
+    // Phase 4: cross-relay direct-call dispatcher task.
+    //
+    // Reads unwrapped (inner, origin_relay_fp) tuples that the
+    // federation layer pushes out of its `handle_signal` arm for
+    // `FederatedSignalForward`, and routes the inner message
+    // through the local signal_hub / call_registry exactly as if
+    // the message had arrived on a local client signal transport.
+    //
+    // In Phase 4 MVP the dispatcher handles:
+    //   * DirectCallOffer  — if target is local, stash in registry
+    //                        with peer_relay_fp and deliver to
+    //                        local callee via signal_hub.
+    //   * DirectCallAnswer — stash callee addr, forward answer to
+    //                        local caller, emit local CallSetup.
+    //   * CallRinging     — forward to local caller for UX.
+    //   * Hangup          — forward to the local participant(s).
+    // Everything else is dropped.
+    {
+        let signal_hub_d = signal_hub.clone();
+        let call_registry_d = call_registry.clone();
+        let advertised_addr_d = advertised_addr_str.clone();
+        let federation_mgr_d = federation_mgr.clone();
+        tokio::spawn(async move {
+            use wzp_proto::{CallAcceptMode, SignalMessage};
+            while let Some((inner, origin_relay_fp)) = cross_relay_rx.recv().await {
+                match inner {
+                    SignalMessage::DirectCallOffer {
+                        ref target_fingerprint,
+                        ref caller_fingerprint,
+                        ref call_id,
+                        ref caller_reflexive_addr,
+                        ref caller_local_addrs,
+                        ..
+                    } => {
+                        // Is the target on THIS relay? If not, drop —
+                        // Phase 4 MVP is single-hop federation only.
+                        let online = {
+                            let hub = signal_hub_d.lock().await;
+                            hub.is_online(target_fingerprint)
+                        };
+                        if !online {
+                            tracing::debug!(
+                                target = %target_fingerprint,
+                                %origin_relay_fp,
+                                "cross-relay: offer target not local, dropping (no multi-hop)"
+                            );
+                            continue;
+                        }
+                        // Stash in local registry so the answer path
+                        // can find the call + route the reply back
+                        // through the same federation link. Include
+                        // Phase 5.5 LAN host candidates too.
+                        {
+                            let mut reg = call_registry_d.lock().await;
+                            reg.create_call(
+                                call_id.clone(),
+                                caller_fingerprint.clone(),
+                                target_fingerprint.clone(),
+                            );
+                            reg.set_caller_reflexive_addr(call_id, caller_reflexive_addr.clone());
+                            reg.set_caller_local_addrs(call_id, caller_local_addrs.clone());
+                            reg.set_peer_relay_fp(call_id, Some(origin_relay_fp.clone()));
+                        }
+                        // Deliver the offer to the local target.
+                        let hub = signal_hub_d.lock().await;
+                        if let Err(e) = hub.send_to(target_fingerprint, &inner).await {
+                            tracing::warn!(
+                                target = %target_fingerprint,
+                                error = %e,
+                                "cross-relay: failed to deliver forwarded offer"
+                            );
+                        }
+                    }
+
+                    SignalMessage::DirectCallAnswer {
+                        ref call_id,
+                        accept_mode,
+                        ref callee_reflexive_addr,
+                        ref callee_local_addrs,
+                        ..
+                    } => {
+                        // Look up the local caller fp from the registry.
+                        let caller_fp = {
+                            let reg = call_registry_d.lock().await;
+                            reg.get(call_id).map(|c| c.caller_fingerprint.clone())
+                        };
+                        let Some(caller_fp) = caller_fp else {
+                            tracing::debug!(%call_id, "cross-relay: answer for unknown call, dropping");
+                            continue;
+                        };
+
+                        if accept_mode == CallAcceptMode::Reject {
+                            // Forward hangup to local caller + clean up registry.
+                            let hub = signal_hub_d.lock().await;
+                            let _ = hub
+                                .send_to(
+                                    &caller_fp,
+                                    &SignalMessage::Hangup {
+                                        reason: wzp_proto::HangupReason::Normal,
+                                        call_id: None,
+                                    },
+                                )
+                                .await;
+                            drop(hub);
+                            let mut reg = call_registry_d.lock().await;
+                            reg.end_call(call_id);
+                            continue;
+                        }
+
+                        // Accept — stash the callee's reflex addr + LAN
+                        // host candidates + mark the call active,
+                        // then read back everything needed to cross-
+                        // wire peer_direct_addr + peer_local_addrs in
+                        // the local CallSetup.
+                        // Also set peer_relay_fp so the originating
+                        // relay knows where to forward MediaPathReport.
+                        let room_name = format!("call-{call_id}");
+                        let (callee_addr_for_setup, callee_local_for_setup) = {
+                            let mut reg = call_registry_d.lock().await;
+                            reg.set_active(call_id, accept_mode, room_name.clone());
+                            reg.set_peer_relay_fp(call_id, Some(origin_relay_fp.clone()));
+                            reg.set_callee_reflexive_addr(
+                                call_id,
+                                callee_reflexive_addr.clone(),
+                            );
+                            reg.set_callee_local_addrs(call_id, callee_local_addrs.clone());
+                            let c = reg.get(call_id);
+                            (
+                                c.and_then(|c| c.callee_reflexive_addr.clone()),
+                                c.map(|c| c.callee_local_addrs.clone()).unwrap_or_default(),
+                            )
+                        };
+
+                        // Forward the raw answer to the local caller so
+                        // the JS side sees DirectCallAnswer (fires any
+                        // "call answered" UX that looks at this message).
+                        {
+                            let hub = signal_hub_d.lock().await;
+                            let _ = hub.send_to(&caller_fp, &inner).await;
+                        }
+
+                        // Emit the LOCAL CallSetup to our local caller.
+                        // relay_addr = our own advertised addr so if P2P
+                        // fails the caller will at least dial OUR relay
+                        // (single-relay fallback — Phase 4.1 will wire
+                        // federated media so that actually reaches the
+                        // peer). peer_direct_addr = the callee's reflex
+                        // addr carried in the answer. peer_local_addrs
+                        // = callee's LAN host candidates (Phase 5.5 ICE).
+                        let setup = SignalMessage::CallSetup {
+                            call_id: call_id.clone(),
+                            room: room_name.clone(),
+                            relay_addr: advertised_addr_d.clone(),
+                            peer_direct_addr: callee_addr_for_setup,
+                            peer_local_addrs: callee_local_for_setup,
+                        };
+                        let hub = signal_hub_d.lock().await;
+                        let _ = hub.send_to(&caller_fp, &setup).await;
+
+                        tracing::info!(
+                            %call_id,
+                            %caller_fp,
+                            %origin_relay_fp,
+                            "cross-relay: delivered answer + CallSetup to local caller"
+                        );
+                    }
+
+                    SignalMessage::CallRinging { ref call_id } => {
+                        // Forward to local caller for "ringing..." UX.
+                        let caller_fp = {
+                            let reg = call_registry_d.lock().await;
+                            reg.get(call_id).map(|c| c.caller_fingerprint.clone())
+                        };
+                        if let Some(fp) = caller_fp {
+                            let hub = signal_hub_d.lock().await;
+                            let _ = hub.send_to(&fp, &inner).await;
+                        }
+                    }
+
+                    // Phase 6: MediaPathReport forwarded across
+                    // federation — deliver to the LOCAL participant.
+                    // The report comes from the remote side, so we
+                    // deliver to whichever participant is local. In
+                    // the cross-relay case, one is local and one is
+                    // remote. Try both — send_to is a no-op if the
+                    // target isn't connected to this relay.
+                    SignalMessage::MediaPathReport { ref call_id, .. } => {
+                        let (caller_fp, callee_fp) = {
+                            let reg = call_registry_d.lock().await;
+                            match reg.get(call_id) {
+                                Some(c) => (
+                                    Some(c.caller_fingerprint.clone()),
+                                    Some(c.callee_fingerprint.clone()),
+                                ),
+                                None => (None, None),
+                            }
+                        };
+                        let hub = signal_hub_d.lock().await;
+                        if let Some(fp) = caller_fp {
+                            let _ = hub.send_to(&fp, &inner).await;
+                        }
+                        if let Some(fp) = callee_fp {
+                            let _ = hub.send_to(&fp, &inner).await;
+                        }
+                    }
+
+                    SignalMessage::Hangup { .. } => {
+                        // Best-effort: broadcast the hangup to every
+                        // local participant of any call that currently
+                        // has this origin as its peer_relay_fp.
+                        // The forwarded hangup doesn't carry a call_id
+                        // so we can't target precisely — Phase 4.1 will
+                        // tighten this once hangup tracking is stricter.
+                        tracing::debug!(
+                            %origin_relay_fp,
+                            "cross-relay: forwarded Hangup (Phase 4.1 will target by call_id)"
+                        );
+                    }
+
+                    _ => {
+                        tracing::debug!(
+                            %origin_relay_fp,
+                            "cross-relay: dispatcher ignoring unsupported inner variant"
+                        );
+                    }
+                }
+            }
+            // Suppress the warning if federation_mgr_d is unused —
+            // it's held here so the Arc doesn't drop during the
+            // dispatcher's lifetime.
+            drop(federation_mgr_d);
+        });
+    }
+
     info!("Listening for connections...");
 
     loop {
-        let connection = match wzp_transport::accept(&endpoint).await {
-            Ok(conn) => conn,
-            Err(e) => { error!("accept: {e}"); continue; }
+        // Pull the next Incoming off the queue. Deliberately do NOT await
+        // the QUIC handshake here — move that into the per-connection
+        // spawned task below. Previously we used wzp_transport::accept
+        // which did both, which meant a single slow handshake would block
+        // the entire accept loop and prevent ALL subsequent connections
+        // from being processed. Surfaced as direct-call hangs where the
+        // callee's call-* connection never completes its QUIC handshake.
+        let incoming = match endpoint.accept().await {
+            Some(inc) => inc,
+            None => {
+                error!("endpoint.accept() returned None — endpoint closed");
+                break;
+            }
         };
 
         let remote_transport = remote_transport.clone();
@@ -493,9 +777,26 @@ async fn main() -> anyhow::Result<()> {
         let federation_mgr = federation_mgr.clone();
         let signal_hub = signal_hub.clone();
         let call_registry = call_registry.clone();
-        let listen_addr_str = config.listen_addr.to_string();
+        let advertised_addr_str = advertised_addr_str.clone();
+        // Phase 4: per-task clone of this relay's federation TLS
+        // fingerprint so the FederatedSignalForward envelopes the
+        // spawned signal handler builds carry `origin_relay_fp`.
+        let tls_fp = tls_fp.clone();
+
+        let incoming_addr = incoming.remote_address();
+        info!(%incoming_addr, "accept queue: new Incoming, spawning handshake task");
 
         tokio::spawn(async move {
+            // Drive the QUIC handshake inside the spawned task so that
+            // slow or hung handshakes never block the outer accept loop.
+            let connection = match incoming.await {
+                Ok(c) => c,
+                Err(e) => {
+                    error!(%incoming_addr, "QUIC handshake failed: {e}");
+                    return;
+                }
+            };
+            info!(%incoming_addr, "QUIC handshake complete");
             let addr = connection.remote_address();
 
             let room_name = connection
@@ -709,6 +1010,7 @@ async fn main() -> anyhow::Result<()> {
                 let _ = transport.send_signal(&SignalMessage::RegisterPresenceAck {
                     success: true,
                     error: None,
+                    relay_build: Some(BUILD_GIT_HASH.to_string()),
                 }).await;
 
                 info!(%addr, fingerprint = %client_fp, alias = ?client_alias, "signal client registered");
@@ -718,9 +1020,17 @@ async fn main() -> anyhow::Result<()> {
                     match transport.recv_signal().await {
                         Ok(Some(msg)) => {
                             match msg {
-                                SignalMessage::DirectCallOffer { ref target_fingerprint, ref call_id, ref caller_alias, .. } => {
+                                SignalMessage::DirectCallOffer {
+                                    ref target_fingerprint,
+                                    ref call_id,
+                                    ref caller_reflexive_addr,
+                                    ref caller_local_addrs,
+                                    ..
+                                } => {
                                     let target_fp = target_fingerprint.clone();
                                     let call_id = call_id.clone();
+                                    let caller_addr_for_registry = caller_reflexive_addr.clone();
+                                    let caller_local_for_registry = caller_local_addrs.clone();
 
                                     // Check if target is online
                                     let online = {
@@ -728,17 +1038,92 @@ async fn main() -> anyhow::Result<()> {
                                         hub.is_online(&target_fp)
                                     };
                                     if !online {
-                                        info!(%addr, target = %target_fp, "call target not online");
+                                        // Phase 4: maybe the target is on a
+                                        // federation peer. Wrap the offer in
+                                        // FederatedSignalForward and broadcast
+                                        // it over every active peer link —
+                                        // whichever relay has the target will
+                                        // unwrap and dispatch locally. We also
+                                        // stash the call in OUR registry so
+                                        // the eventual answer coming back via
+                                        // federation has a matching entry.
+                                        let forwarded = if let Some(ref fm) = federation_mgr {
+                                            let forward = SignalMessage::FederatedSignalForward {
+                                                inner: Box::new(msg.clone()),
+                                                origin_relay_fp: tls_fp.clone(),
+                                            };
+                                            let count = fm.broadcast_signal(&forward).await;
+                                            if count > 0 {
+                                                info!(
+                                                    %addr,
+                                                    target = %target_fp,
+                                                    peers = count,
+                                                    "direct-call offer forwarded to federation peers"
+                                                );
+                                                true
+                                            } else {
+                                                false
+                                            }
+                                        } else {
+                                            false
+                                        };
+
+                                        if !forwarded {
+                                            info!(%addr, target = %target_fp, "call target not online (no federation route)");
                                             let _ = transport.send_signal(&SignalMessage::Hangup {
                                                 reason: wzp_proto::HangupReason::Normal,
+                                                call_id: None,
                                             }).await;
                                             continue;
                                         }
 
-                                    // Create call in registry
+                                        // Create call in registry with the
+                                        // caller's reflex addr + LAN host
+                                        // candidates, and mark it as
+                                        // cross-relay so the answer path knows
+                                        // to route the CallSetup's
+                                        // peer_direct_addr from what the
+                                        // federated answer carries. peer_relay_fp
+                                        // stays None here because we broadcast —
+                                        // the receiving relay picks itself as
+                                        // the answer source and its forwarded
+                                        // answer will identify itself there.
+                                        {
+                                            let mut reg = call_registry.lock().await;
+                                            reg.create_call(
+                                                call_id.clone(),
+                                                client_fp.clone(),
+                                                target_fp.clone(),
+                                            );
+                                            reg.set_caller_reflexive_addr(
+                                                &call_id,
+                                                caller_addr_for_registry.clone(),
+                                            );
+                                            reg.set_caller_local_addrs(
+                                                &call_id,
+                                                caller_local_for_registry.clone(),
+                                            );
+                                        }
+
+                                        // Send ringing to caller immediately
+                                        // so the UI shows feedback while the
+                                        // federated delivery is in flight.
+                                        let _ = transport.send_signal(&SignalMessage::CallRinging {
+                                            call_id: call_id.clone(),
+                                        }).await;
+                                        continue;
+                                    }
+
+                                    // Create call in registry + stash the caller's
+                                    // reflex addr (Phase 3 hole-punching) AND its
+                                    // LAN host candidates (Phase 5.5 ICE). The
+                                    // relay treats both as opaque. Both are
+                                    // injected later into the callee's CallSetup.
                                     {
                                         let mut reg = call_registry.lock().await;
                                         reg.create_call(call_id.clone(), client_fp.clone(), target_fp.clone());
+                                        reg.set_caller_reflexive_addr(&call_id, caller_addr_for_registry);
+                                        reg.set_caller_local_addrs(&call_id, caller_local_for_registry);
                                     }
 
                                     // Forward offer to callee
@@ -755,16 +1140,37 @@ async fn main() -> anyhow::Result<()> {
                                     }).await;
                                 }
 
-                                SignalMessage::DirectCallAnswer { ref call_id, ref accept_mode, .. } => {
+                                SignalMessage::DirectCallAnswer {
+                                    ref call_id,
+                                    ref accept_mode,
+                                    ref callee_reflexive_addr,
+                                    ref callee_local_addrs,
+                                    ..
+                                } => {
                                     let call_id = call_id.clone();
                                     let mode = *accept_mode;
+                                    let callee_addr_for_registry = callee_reflexive_addr.clone();
+                                    let callee_local_for_registry = callee_local_addrs.clone();
 
-                                    let peer_fp = {
+                                    // Phase 4: look up peer fingerprint AND
+                                    // peer_relay_fp in one lock acquisition.
+                                    // peer_relay_fp being Some means the
+                                    // caller is on a remote federation peer
+                                    // and we have to route the answer /
+                                    // hangup back through that link instead
+                                    // of local signal_hub.
+                                    let (peer_fp, peer_relay_fp) = {
                                         let reg = call_registry.lock().await;
-                                        reg.peer_fingerprint(&call_id, &client_fp).map(|s| s.to_string())
+                                        match reg.get(&call_id) {
+                                            Some(c) => (
+                                                Some(reg.peer_fingerprint(&call_id, &client_fp).map(|s| s.to_string())),
+                                                c.peer_relay_fp.clone(),
+                                            ),
+                                            None => (None, None),
+                                        }
                                     };
 
-                                    let Some(peer_fp) = peer_fp else {
+                                    let Some(Some(peer_fp)) = peer_fp else {
                                         warn!(call_id = %call_id, "answer for unknown call");
                                         continue;
                                     };
@@ -774,58 +1180,151 @@ async fn main() -> anyhow::Result<()> {
                                         let mut reg = call_registry.lock().await;
                                         reg.end_call(&call_id);
                                         drop(reg);
+
+                                        // Phase 4: cross-relay reject —
+                                        // forward the hangup to the origin
+                                        // relay instead of local signal_hub.
+                                        if let Some(ref origin_fp) = peer_relay_fp {
+                                            if let Some(ref fm) = federation_mgr {
+                                                let hangup = SignalMessage::Hangup {
+                                                    reason: wzp_proto::HangupReason::Normal,
+                                                    call_id: Some(call_id.clone()),
+                                                };
+                                                let forward = SignalMessage::FederatedSignalForward {
+                                                    inner: Box::new(hangup),
+                                                    origin_relay_fp: tls_fp.clone(),
+                                                };
+                                                if let Err(e) = fm.send_signal_to_peer(origin_fp, &forward).await {
+                                                    warn!(%call_id, %origin_fp, error = %e, "cross-relay reject forward failed");
+                                                }
+                                            }
+                                        } else {
                                             let hub = signal_hub.lock().await;
                                             let _ = hub.send_to(&peer_fp, &SignalMessage::Hangup {
                                                 reason: wzp_proto::HangupReason::Normal,
+                                                call_id: Some(call_id.clone()),
                                             }).await;
+                                        }
                                     } else {
-                                        // Accept — create private room
+                                        // Accept — create private room + stash the
+                                        // callee's reflex addr if it advertised one
+                                        // (AcceptTrusted only — privacy-mode answers
+                                        // leave it None by design). Then read back
+                                        // BOTH parties' addrs so we can cross-wire
+                                        // peer_direct_addr on the CallSetups below.
                                         let room = format!("call-{call_id}");
-                                        {
+                                        let (caller_addr, callee_addr, caller_local, callee_local) = {
                                             let mut reg = call_registry.lock().await;
                                             reg.set_active(&call_id, mode, room.clone());
-                                        }
-                                        info!(call_id = %call_id, room = %room, mode = ?mode, "call accepted, creating room");
+                                            reg.set_callee_reflexive_addr(&call_id, callee_addr_for_registry);
+                                            reg.set_callee_local_addrs(&call_id, callee_local_for_registry.clone());
+                                            let call = reg.get(&call_id);
+                                            (
+                                                call.and_then(|c| c.caller_reflexive_addr.clone()),
+                                                call.and_then(|c| c.callee_reflexive_addr.clone()),
+                                                call.map(|c| c.caller_local_addrs.clone()).unwrap_or_default(),
+                                                call.map(|c| c.callee_local_addrs.clone()).unwrap_or_default(),
+                                            )
+                                        };
+                                        info!(
+                                            call_id = %call_id,
+                                            room = %room,
+                                            ?mode,
+                                            p2p_viable = caller_addr.is_some() && callee_addr.is_some(),
+                                            "call accepted, creating room"
+                                        );
 
+                                        let relay_addr_for_setup = advertised_addr_str.clone();
+
+                                        if let Some(ref origin_fp) = peer_relay_fp {
+                                            // Phase 4 cross-relay: the caller
+                                            // is on a remote peer. Forward the
+                                            // raw answer (which carries the
+                                            // callee's reflex addr) back over
+                                            // federation — the peer's
+                                            // cross-relay dispatcher will
+                                            // deliver it to the local caller
+                                            // AND emit a CallSetup on that
+                                            // side with peer_direct_addr =
+                                            // callee_addr.
+                                            //
+                                            // Here we emit only the LOCAL
+                                            // CallSetup (to our callee) with
+                                            // peer_direct_addr = caller_addr.
+                                            if let Some(ref fm) = federation_mgr {
+                                                let forward = SignalMessage::FederatedSignalForward {
+                                                    inner: Box::new(msg.clone()),
+                                                    origin_relay_fp: tls_fp.clone(),
+                                                };
+                                                if let Err(e) = fm.send_signal_to_peer(origin_fp, &forward).await {
+                                                    warn!(
+                                                        %call_id,
+                                                        %origin_fp,
+                                                        error = %e,
+                                                        "cross-relay answer forward failed"
+                                                    );
+                                                }
+                                            }
+
+                                            let setup_for_callee = SignalMessage::CallSetup {
+                                                call_id: call_id.clone(),
+                                                room: room.clone(),
+                                                relay_addr: relay_addr_for_setup,
+                                                peer_direct_addr: caller_addr.clone(),
+                                                peer_local_addrs: caller_local.clone(),
+                                            };
+                                            let hub = signal_hub.lock().await;
+                                            let _ = hub.send_to(&client_fp, &setup_for_callee).await;
+                                        } else {
+                                            // Local call (existing Phase 3 path).
                                             // Forward answer to caller
                                             {
                                                 let hub = signal_hub.lock().await;
                                                 let _ = hub.send_to(&peer_fp, &msg).await;
                                             }
 
-                                        // Send CallSetup to both parties
-                                        // Use the address the client connected to (their remote addr
-                                        // is our perspective, but we need our listen addr).
-                                        // Replace 0.0.0.0 with the client's destination IP.
-                                        let relay_addr_for_setup = if listen_addr_str.starts_with("0.0.0.0:") {
-                                            let port = &listen_addr_str[8..];
-                                            // Use the local IP from the client's connection
-                                            let local_ip = addr.ip();
-                                            if local_ip.is_loopback() {
-                                                format!("127.0.0.1:{port}")
-                                            } else {
-                                                format!("{local_ip}:{port}")
-                                            }
-                                        } else {
-                                            listen_addr_str.clone()
+                                            // Send CallSetup to BOTH parties with
+                                            // cross-wired peer_direct_addr +
+                                            // peer_local_addrs (Phase 5.5 ICE).
+                                            let setup_for_caller = SignalMessage::CallSetup {
+                                                call_id: call_id.clone(),
+                                                room: room.clone(),
+                                                relay_addr: relay_addr_for_setup.clone(),
+                                                peer_direct_addr: callee_addr.clone(),
+                                                peer_local_addrs: callee_local.clone(),
                                             };
-                                        let setup = SignalMessage::CallSetup {
+                                            let setup_for_callee = SignalMessage::CallSetup {
                                                 call_id: call_id.clone(),
                                                 room: room.clone(),
                                                 relay_addr: relay_addr_for_setup,
+                                                peer_direct_addr: caller_addr.clone(),
+                                                peer_local_addrs: caller_local.clone(),
                                             };
-                                        {
                                             let hub = signal_hub.lock().await;
-                                            let _ = hub.send_to(&peer_fp, &setup).await;
-                                            let _ = hub.send_to(&client_fp, &setup).await;
+                                            let _ = hub.send_to(&peer_fp, &setup_for_caller).await;
+                                            let _ = hub.send_to(&client_fp, &setup_for_callee).await;
                                         }
                                     }
                                 }
 
-                                SignalMessage::Hangup { .. } => {
-                                    // Forward hangup to all active calls for this user
+                                SignalMessage::Hangup { ref call_id, .. } => {
+                                    // If the client sent a call_id, only end
+                                    // that specific call. Otherwise (old clients)
+                                    // fall back to ending ALL active calls for
+                                    // this user — which can race with new calls.
                                     let calls = {
                                         let reg = call_registry.lock().await;
+                                        if let Some(cid) = call_id {
+                                            // Targeted hangup: only the named call
+                                            reg.get(cid)
+                                                .map(|c| vec![(c.call_id.clone(), if c.caller_fingerprint == client_fp {
+                                                    c.callee_fingerprint.clone()
+                                                } else {
+                                                    c.caller_fingerprint.clone()
+                                                })])
+                                                .unwrap_or_default()
+                                        } else {
+                                            // Legacy: end all calls for this user
                                             reg.calls_for_fingerprint(&client_fp)
                                                 .iter()
                                                 .map(|c| (c.call_id.clone(), if c.caller_fingerprint == client_fp {
@@ -834,13 +1333,58 @@ async fn main() -> anyhow::Result<()> {
                                                     c.caller_fingerprint.clone()
                                                 }))
                                                 .collect::<Vec<_>>()
+                                        }
                                     };
-                                    for (call_id, peer_fp) in &calls {
+                                    for (cid, peer_fp) in &calls {
                                         let hub = signal_hub.lock().await;
                                         let _ = hub.send_to(peer_fp, &msg).await;
                                         drop(hub);
                                         let mut reg = call_registry.lock().await;
-                                        reg.end_call(call_id);
+                                        reg.end_call(cid);
+                                    }
+                                }
+
+                                // Phase 6: forward MediaPathReport to the
+                                // call peer so both sides can negotiate
+                                // the media path before committing.
+                                SignalMessage::MediaPathReport { ref call_id, .. } => {
+                                    // Look up peer AND check if this is a
+                                    // cross-relay call (same pattern as
+                                    // DirectCallAnswer).
+                                    let (peer_fp, peer_relay_fp) = {
+                                        let reg = call_registry.lock().await;
+                                        match reg.get(call_id) {
+                                            Some(c) => (
+                                                reg.peer_fingerprint(call_id, &client_fp)
+                                                    .map(|s| s.to_string()),
+                                                c.peer_relay_fp.clone(),
+                                            ),
+                                            None => (None, None),
+                                        }
+                                    };
+
+                                    if let Some(fp) = peer_fp {
+                                        if let Some(ref origin_fp) = peer_relay_fp {
+                                            // Cross-relay: wrap and forward
+                                            if let Some(ref fm) = federation_mgr {
+                                                let forward = SignalMessage::FederatedSignalForward {
+                                                    inner: Box::new(msg.clone()),
+                                                    origin_relay_fp: tls_fp.clone(),
+                                                };
+                                                if let Err(e) = fm.send_signal_to_peer(origin_fp, &forward).await {
+                                                    warn!(
+                                                        %call_id,
+                                                        %origin_fp,
+                                                        error = %e,
+                                                        "cross-relay MediaPathReport forward failed"
+                                                    );
+                                                }
+                                            }
+                                        } else {
+                                            // Local call
+                                            let hub = signal_hub.lock().await;
+                                            let _ = hub.send_to(&fp, &msg).await;
+                                        }
                                     }
                                 }
 
@@ -848,6 +1392,31 @@ async fn main() -> anyhow::Result<()> {
                                     let _ = transport.send_signal(&SignalMessage::Pong { timestamp_ms }).await;
                                 }
 
+                                // QUIC-native NAT reflection ("STUN for QUIC").
+                                // The client asks "what source address do you
+                                // see for me?" and we reply with whatever
+                                // quinn reports as this connection's remote
+                                // address — i.e. the post-NAT public address
+                                // as observed from the server side of the TLS
+                                // session. Used by the P2P path to learn the
+                                // client's server-reflexive address without
+                                // running a separate STUN server. No auth or
+                                // rate-limit in Phase 1 — the client is
+                                // already TLS-authenticated by the time it
+                                // reaches this match arm.
+                                SignalMessage::Reflect => {
+                                    let observed_addr = addr.to_string();
+                                    if let Err(e) = transport.send_signal(
+                                        &SignalMessage::ReflectResponse {
+                                            observed_addr: observed_addr.clone(),
+                                        },
+                                    ).await {
+                                        warn!(%addr, error = %e, "reflect: failed to send response");
+                                    } else {
+                                        debug!(%addr, %observed_addr, "reflect: responded");
+                                    }
+                                }
+
                                 other => {
                                     warn!(%addr, "signal: unexpected message: {:?}", std::mem::discriminant(&other));
                                 }
@@ -857,6 +1426,16 @@ async fn main() -> anyhow::Result<()> {
                             info!(%addr, "signal connection closed");
                             break;
                         }
+                        Err(wzp_proto::TransportError::Deserialize(e)) => {
+                            // Forward-compat: the peer sent a
+                            // SignalMessage variant we don't know
+                            // (newer client, newer federation peer).
+                            // Log and continue — tearing down the
+                            // connection on unknown variants would
+                            // silently kill interop across minor
+                            // protocol version bumps.
+                            warn!(%addr, "signal deserialize (unknown variant?), continuing: {e}");
+                        }
                         Err(e) => {
                             warn!(%addr, "signal recv error: {e}");
                             break;
@@ -880,6 +1459,7 @@ async fn main() -> anyhow::Result<()> {
                     let hub = signal_hub.lock().await;
                     let _ = hub.send_to(peer_fp, &SignalMessage::Hangup {
                         reason: wzp_proto::HangupReason::Normal,
+                        call_id: Some(call_id.clone()),
                     }).await;
                     drop(hub);
                     let mut reg = call_registry.lock().await;
@@ -1153,4 +1733,5 @@ async fn main() -> anyhow::Result<()> {
             }
         });
     }
+    Ok(())
 }

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

@@ -10,9 +10,11 @@ use std::time::Duration;
 
 use bytes::Bytes;
 use tokio::sync::Mutex;
-use tracing::{debug, error, info, trace, warn};
+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;
@@ -50,6 +52,45 @@ impl DebugTap {
     }
 }
 
+/// 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_by_key(|t| match t {
+            Tier::Good => 2,
+            Tier::Degraded => 1,
+            Tier::Catastrophic => 0,
+        })
+        .unwrap_or(Tier::Good)
+}
+
 /// Unique participant ID within a room.
 pub type ParticipantId = u64;
 
@@ -208,6 +249,10 @@ pub struct RoomManager {
     acl: Option<HashMap<String, HashSet<String>>>,
     /// Channel for room lifecycle events (federation subscribes).
     event_tx: tokio::sync::broadcast::Sender<RoomEvent>,
+    /// Per-participant quality tracking, keyed by (room_name, participant_id).
+    qualities: HashMap<(String, ParticipantId), ParticipantQuality>,
+    /// Current room-wide tier per room (to avoid repeated broadcasts).
+    room_tiers: HashMap<String, Tier>,
 }
 
 impl RoomManager {
@@ -217,6 +262,8 @@ impl RoomManager {
             rooms: HashMap::new(),
             acl: None,
             event_tx,
+            qualities: HashMap::new(),
+            room_tiers: HashMap::new(),
         }
     }
 
@@ -227,6 +274,8 @@ impl RoomManager {
             rooms: HashMap::new(),
             acl: Some(HashMap::new()),
             event_tx,
+            qualities: HashMap::new(),
+            room_tiers: HashMap::new(),
         }
     }
 
@@ -277,6 +326,7 @@ impl RoomManager {
             || 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 id = room.add(addr, sender, fingerprint.map(|s| s.to_string()), alias.map(|s| s.to_string()));
+        self.qualities.insert((room_name.to_string(), id), ParticipantQuality::new());
         if was_empty {
             let _ = self.event_tx.send(RoomEvent::LocalJoin { room: room_name.to_string() });
         }
@@ -323,10 +373,12 @@ 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>)> {
+        self.qualities.remove(&(room_name.to_string(), participant_id));
         if let Some(room) = self.rooms.get_mut(room_name) {
             room.remove(participant_id);
             if room.is_empty() {
                 self.rooms.remove(room_name);
+                self.room_tiers.remove(room_name);
                 let _ = self.event_tx.send(RoomEvent::LocalLeave { room: room_name.to_string() });
                 info!(room = room_name, "room closed (empty)");
                 return None;
@@ -363,6 +415,58 @@ impl RoomManager {
     pub fn list(&self) -> Vec<(String, usize)> {
         self.rooms.iter().map(|(k, v)| (k.clone(), v.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(
+        &mut self,
+        room_name: &str,
+        participant_id: ParticipantId,
+        report: &wzp_proto::packet::QualityReport,
+    ) -> Option<(wzp_proto::SignalMessage, Vec<ParticipantSender>)> {
+        let key = (room_name.to_string(), participant_id);
+        let tier_changed = self.qualities
+            .get_mut(&key)
+            .and_then(|pq| pq.observe(report))
+            .is_some();
+
+        if !tier_changed {
+            return None;
+        }
+
+        // Compute the weakest tier across all participants in this room
+        let room_qualities = self.qualities.iter()
+            .filter(|((rn, _), _)| rn == room_name)
+            .map(|(_, pq)| pq);
+        let weakest = weakest_tier(room_qualities);
+
+        let current_room_tier = self.room_tiers.get(room_name).copied().unwrap_or(Tier::Good);
+        if weakest == current_room_tier {
+            return None;
+        }
+
+        // Room-wide tier changed — update and broadcast directive
+        self.room_tiers.insert(room_name.to_string(), weakest);
+        let profile = weakest.profile();
+        info!(
+            room = room_name,
+            old_tier = ?current_room_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 = self.rooms.get(room_name)
+            .map(|r| r.all_senders())
+            .unwrap_or_default();
+        Some((directive, senders))
+    }
 }
 
 // ---------------------------------------------------------------------------
@@ -382,18 +486,32 @@ 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)?;
@@ -483,7 +601,6 @@ 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 +639,17 @@ 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 mut mgr = room_mgr.lock().await;
+            let directive = if let Some(ref report) = pkt.quality_report {
+                mgr.observe_quality(&room_name, participant_id, report)
+            } else {
+                None
+            };
+            let o = mgr.others(&room_name, participant_id);
+            (o, directive)
         };
         let lock_ms = lock_start.elapsed().as_millis() as u64;
         if lock_ms > 10 {
@@ -538,6 +661,11 @@ async fn run_participant_plain(
             );
         }
 
+        // Broadcast quality directive to all participants if tier changed
+        if let Some((directive, all_senders)) = quality_directive {
+            broadcast_signal(&all_senders, &directive).await;
+        }
+
         // Debug tap: log packet metadata
         if let Some(ref tap) = debug_tap {
             if tap.matches(&room_name) {
@@ -706,9 +834,15 @@ 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 mut mgr = room_mgr.lock().await;
+                    let directive = if let Some(ref report) = pkt.quality_report {
+                        mgr.observe_quality(&room_name, participant_id, report)
+                    } else {
+                        None
+                    };
+                    let o = mgr.others(&room_name, participant_id);
+                    (o, directive)
                 };
                 let lock_ms = lock_start.elapsed().as_millis() as u64;
                 if lock_ms > 10 {
@@ -720,6 +854,11 @@ 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 {
@@ -838,7 +977,7 @@ 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());
     }
@@ -960,4 +1099,47 @@ 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;
 
@@ -94,7 +94,7 @@ 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/tests/cross_relay_direct_call.rs

@@ -0,0 +1,317 @@
+//! 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_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_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(),
+    };
+    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(),
+    }
+}
+
+// ────────────────────────────────────────────────────────────────
+// 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/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 +1,294 @@
+//! 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(),
+    };
+    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(),
+    };
+    (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_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_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 +1,229 @@
+//! 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,
+                                })
+                                .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 +1,291 @@
+<!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">
+      <!-- Connect screen -->
+      <div id="connect-screen">
+        <h1>WarzonePhone</h1>
+        <p class="subtitle">Encrypted Voice</p>
+        <div class="form">
+          <label>Relay
+            <button id="relay-selected" class="relay-selected" type="button">
+              <span id="relay-dot" class="dot"></span>
+              <span id="relay-label">Select relay...</span>
+              <span class="arrow">&#9881;</span>
+            </button>
+          </label>
+          <label>Room
+            <input id="room" type="text" value="general" />
+          </label>
+          <label>Alias
+            <input id="alias" type="text" placeholder="your name" />
+          </label>
+          <div class="form-row">
+            <label class="checkbox">
+              <input id="os-aec" type="checkbox" checked />
+              OS Echo Cancel
+            </label>
+            <button id="settings-btn-home" class="icon-btn" title="Settings (Cmd+,)">&#9881;</button>
+          </div>
+          <!-- Mode toggle -->
+          <div class="mode-toggle" style="display:flex;gap:8px;margin-bottom:8px;">
+            <button id="mode-room" class="mode-btn active" style="flex:1">Room</button>
+            <button id="mode-direct" class="mode-btn" style="flex:1">Direct Call</button>
+          </div>
+
+          <!-- Room mode (default) -->
+          <div id="room-mode">
+            <button id="connect-btn" class="primary">Connect</button>
+          </div>
+
+          <!-- Direct call mode -->
+          <div id="direct-mode" class="hidden">
+            <button id="register-btn" class="primary" style="background:#2196F3">Register on Relay</button>
+            <div id="direct-registered" class="hidden" style="margin-top:12px">
+              <div class="direct-registered-header">
+                <p id="registered-status" style="color:var(--green);font-size:13px;margin:0">&#x2705; Registered — waiting for calls</p>
+                <button id="deregister-btn" class="secondary-btn small">Deregister</button>
+              </div>
+              <div id="incoming-call-panel" class="hidden" style="background:#1B5E20;padding:12px;border-radius:8px;margin:8px 0">
+                <p style="font-weight:bold;margin:0 0 4px 0">Incoming Call</p>
+                <p id="incoming-caller" style="font-size:12px;opacity:0.8;margin:0 0 8px 0">From: unknown</p>
+                <div style="display:flex;gap:8px">
+                  <button id="accept-call-btn" style="flex:1;background:var(--green);color:white;border:none;padding:8px;border-radius:6px;cursor:pointer">Accept</button>
+                  <button id="reject-call-btn" style="flex:1;background:var(--red);color:white;border:none;padding:8px;border-radius:6px;cursor:pointer">Reject</button>
+                </div>
+              </div>
+
+              <!-- Recent contacts -->
+              <div id="recent-contacts-section" class="hidden">
+                <div class="history-header">Recent contacts</div>
+                <div id="recent-contacts-list" class="history-list"></div>
+              </div>
+
+              <!-- Call history -->
+              <div id="call-history-section" class="hidden">
+                <div class="history-header">
+                  History
+                  <button id="clear-history-btn" class="link-btn">clear</button>
+                </div>
+                <div id="call-history-list" class="history-list"></div>
+              </div>
+
+              <label style="margin-top:8px">Call by fingerprint
+                <input id="target-fp" type="text" placeholder="xxxx:xxxx:xxxx:..." />
+              </label>
+              <button id="call-btn" class="primary" style="margin-top:8px">Call</button>
+              <p id="call-status-text" style="color:var(--yellow);font-size:13px;margin-top:4px"></p>
+            </div>
+          </div>
+
+          <p id="connect-error" class="error"></p>
+        </div>
+        <div class="identity-info">
+          <span id="my-identicon"></span>
+          <span id="my-fingerprint" class="fp-display"></span>
+        </div>
+        <div class="recent-rooms" id="recent-rooms"></div>
+      </div>
+
+      <!-- In-call screen -->
+      <div id="call-screen" class="hidden">
+        <div class="call-header">
+          <div class="call-header-row">
+            <div id="room-name" class="room-name"></div>
+            <button id="settings-btn-call" class="icon-btn small" title="Settings (Cmd+,)">&#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 — shown instead of the group
+             participant list when directCallPeer is set. Centered
+             identicon, name, fp, connection badge. Hidden for
+             room calls (directCallPeer == null). -->
+        <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>
+        <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>
+
+      <!-- Settings panel -->
+      <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-label">Auto</span>
+              </div>
+              <input id="s-quality" type="range" min="0" max="7" step="1" value="3" class="quality-slider" />
+              <div class="quality-ticks">
+                <span>64k</span>
+                <span>48k</span>
+                <span>32k</span>
+                <span>Auto</span>
+                <span>24k</span>
+                <span>6k</span>
+                <span>C2</span>
+                <span>1.2k</span>
+              </div>
+            </div>
+            <label class="checkbox">
+              <input id="s-os-aec" type="checkbox" />
+              OS Echo Cancellation (macOS VoiceProcessingIO)
+            </label>
+            <label class="checkbox">
+              <input id="s-agc" type="checkbox" checked />
+              Automatic Gain Control
+            </label>
+            <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>
+          </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>
+            <small style="color:var(--text-dim);display:block;margin-top:4px">
+              Rolling buffer of the last 200 call-flow events. Turned off by
+              default — the GUI overlay only populates when the checkbox above
+              is on, but logcat (adb) always keeps a copy regardless.
+            </small>
+          </div>
+          <div class="settings-section">
+            <h3>Identity</h3>
+            <div class="setting-row">
+              <span class="setting-label">Fingerprint</span>
+              <span id="s-fingerprint" class="fp-display-large"></span>
+            </div>
+            <div class="setting-row">
+              <span class="setting-label">Identity file</span>
+              <span class="fp-display">~/.wzp/identity</span>
+            </div>
+          </div>
+          <div class="settings-section">
+            <h3>Network</h3>
+            <div class="setting-row">
+              <span class="setting-label">Public address</span>
+              <span id="s-reflected-addr" class="fp-display">(not queried)</span>
+              <button id="s-reflect-btn" class="secondary-btn">Detect</button>
+            </div>
+            <small style="color:var(--text-dim);display:block;margin-top:4px">
+              Asks the registered relay to echo back the IP:port it sees for this
+              connection (QUIC-native NAT reflection, replaces STUN).
+            </small>
+            <div class="setting-row" style="margin-top:10px">
+              <span class="setting-label">NAT type</span>
+              <span id="s-nat-type" class="fp-display">(not detected)</span>
+              <button id="s-nat-detect-btn" class="secondary-btn">Detect NAT</button>
+            </div>
+            <div id="s-nat-probes" style="margin-top:6px;font-size:11px;color:var(--text-dim)"></div>
+            <small style="color:var(--text-dim);display:block;margin-top:4px">
+              Probes every configured relay in parallel and compares the results
+              to classify the NAT: cone (P2P viable), symmetric (must relay),
+              multiple, or unknown.
+            </small>
+          </div>
+          <div class="settings-section">
+            <h3>Recent Rooms</h3>
+            <div id="s-recent-rooms" class="recent-rooms-list"></div>
+            <button id="s-clear-recent" class="secondary-btn">Clear History</button>
+          </div>
+          <button id="settings-save" class="primary">Save</button>
+        </div>
+      </div>
+
+      <!-- Manage Relays dialog -->
+      <div id="relay-dialog" class="hidden">
+        <div class="settings-card relay-dialog-card">
+          <div class="settings-header">
+            <h2>Manage Relays</h2>
+            <button id="relay-dialog-close" class="icon-btn">&times;</button>
+          </div>
+          <div id="relay-dialog-list" class="relay-dialog-list"></div>
+          <div class="relay-add-row">
+            <div class="relay-add-inputs">
+              <input id="relay-add-name" type="text" placeholder="Name" />
+              <input id="relay-add-addr" type="text" placeholder="host:port" />
+            </div>
+            <button id="relay-add-btn" class="primary">Add Relay</button>
+          </div>
+        </div>
+      </div>
+      <!-- Key changed warning dialog -->
+      <div id="key-warning" class="hidden">
+        <div class="settings-card key-warning-card">
+          <div class="key-warning-icon">&#9888;</div>
+          <h2>Server Key Changed</h2>
+          <p class="key-warning-text">The relay's identity has changed since you last connected. This usually happens when the server was restarted, but could also indicate a security issue.</p>
+          <div class="key-warning-fps">
+            <div class="key-fp-row">
+              <span class="key-fp-label">Previously known</span>
+              <code id="kw-old-fp" class="key-fp"></code>
+            </div>
+            <div class="key-fp-row">
+              <span class="key-fp-label">New key</span>
+              <code id="kw-new-fp" class="key-fp"></code>
+            </div>
+          </div>
+          <div class="key-warning-actions">
+            <button id="kw-accept" class="primary">Accept New Key</button>
+            <button id="kw-cancel" class="secondary-btn">Cancel</button>
+          </div>
+        </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)
+}

desktop/src-tauri/src/history.rs

@@ -0,0 +1,180 @@
+//! Call history store.
+//!
+//! Keeps a rolling JSON file of the last N direct-call events so the UI can
+//! show "recent contacts" + "call history with callback buttons" on the
+//! direct-call screen. Storage lives in `<APP_DATA_DIR>/call_history.json`
+//! alongside the identity file. The file is read lazily on first access and
+//! cached in an RwLock behind a OnceLock.
+//!
+//! This is a v1 — no duration tracking yet, entries are logged at the
+//! moment the direction is decided (placed / received / missed).
+
+use std::path::PathBuf;
+use std::sync::{OnceLock, RwLock};
+use std::time::{SystemTime, UNIX_EPOCH};
+
+use serde::{Deserialize, Serialize};
+
+/// Maximum number of history entries we keep. Older ones are pruned FIFO.
+const MAX_ENTRIES: usize = 200;
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(rename_all = "lowercase")]
+pub enum CallDirection {
+    /// Local user placed the call.
+    Placed,
+    /// Remote user called and local user answered.
+    Received,
+    /// Remote user called but local user did not answer (rejected or
+    /// missed entirely — the UI treats these identically).
+    Missed,
+}
+
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct CallHistoryEntry {
+    pub call_id: String,
+    pub peer_fp: String,
+    pub peer_alias: Option<String>,
+    pub direction: CallDirection,
+    /// Seconds since UNIX epoch, UTC.
+    pub timestamp_unix: u64,
+}
+
+// ─── In-process store (loaded from disk once) ─────────────────────────────
+
+static STORE: OnceLock<RwLock<Vec<CallHistoryEntry>>> = OnceLock::new();
+
+fn store() -> &'static RwLock<Vec<CallHistoryEntry>> {
+    STORE.get_or_init(|| RwLock::new(load_from_disk()))
+}
+
+fn history_path() -> PathBuf {
+    crate::APP_DATA_DIR
+        .get()
+        .cloned()
+        .unwrap_or_else(|| {
+            let home = std::env::var("HOME").unwrap_or_else(|_| ".".into());
+            PathBuf::from(home).join(".wzp")
+        })
+        .join("call_history.json")
+}
+
+fn load_from_disk() -> Vec<CallHistoryEntry> {
+    let path = history_path();
+    let Ok(bytes) = std::fs::read(&path) else {
+        return Vec::new();
+    };
+    serde_json::from_slice::<Vec<CallHistoryEntry>>(&bytes)
+        .inspect_err(|e| tracing::warn!(path = %path.display(), error = %e, "call_history.json parse failed"))
+        .unwrap_or_default()
+}
+
+fn save_to_disk(entries: &[CallHistoryEntry]) {
+    let path = history_path();
+    if let Some(parent) = path.parent() {
+        let _ = std::fs::create_dir_all(parent);
+    }
+    let Ok(json) = serde_json::to_vec_pretty(entries) else { return };
+    // Atomic write via temp file + rename so a crash mid-write doesn't
+    // leave us with a half-file on disk.
+    let tmp = path.with_extension("json.tmp");
+    if std::fs::write(&tmp, &json).is_ok() {
+        let _ = std::fs::rename(&tmp, &path);
+    }
+}
+
+fn now_unix() -> u64 {
+    SystemTime::now()
+        .duration_since(UNIX_EPOCH)
+        .map(|d| d.as_secs())
+        .unwrap_or(0)
+}
+
+// ─── Public API ───────────────────────────────────────────────────────────
+
+/// Append a new entry to the store and persist to disk. Trims the store to
+/// `MAX_ENTRIES` after insertion.
+pub fn log(
+    call_id: String,
+    peer_fp: String,
+    peer_alias: Option<String>,
+    direction: CallDirection,
+) {
+    tracing::info!(
+        %call_id, %peer_fp, ?direction,
+        alias = ?peer_alias,
+        "history::log"
+    );
+    let entry = CallHistoryEntry {
+        call_id: call_id.clone(),
+        peer_fp,
+        peer_alias,
+        direction,
+        timestamp_unix: now_unix(),
+    };
+    let mut guard = store().write().unwrap();
+    // If an entry for this call_id already exists, update it in-place
+    // rather than appending a duplicate. Protects against the caller
+    // side adding a second Missed row when the callee's DirectCallOffer
+    // bounces back through federation / loopback, or when some future
+    // relay routing edge case double-emits a signal. The dedup keeps
+    // history tidy and matches what the user intuitively expects (one
+    // history row per call, not one per signal event).
+    if let Some(existing) = guard.iter_mut().rev().find(|e| e.call_id == call_id) {
+        tracing::info!(%call_id, from = ?existing.direction, to = ?direction, "history::log replacing existing entry");
+        existing.direction = direction;
+        existing.timestamp_unix = entry.timestamp_unix;
+        save_to_disk(&guard);
+        return;
+    }
+    guard.push(entry);
+    if guard.len() > MAX_ENTRIES {
+        let drop_n = guard.len() - MAX_ENTRIES;
+        guard.drain(0..drop_n);
+    }
+    save_to_disk(&guard);
+}
+
+/// Return a copy of all entries in reverse-chronological order
+/// (most recent first).
+pub fn all() -> Vec<CallHistoryEntry> {
+    let guard = store().read().unwrap();
+    guard.iter().rev().cloned().collect()
+}
+
+/// Unique peer contacts sorted by most recent interaction. Each contact
+/// is represented by the newest history entry for that fingerprint.
+pub fn contacts() -> Vec<CallHistoryEntry> {
+    let guard = store().read().unwrap();
+    let mut seen: std::collections::HashSet<String> = std::collections::HashSet::new();
+    let mut out = Vec::new();
+    // iterate newest → oldest
+    for entry in guard.iter().rev() {
+        if seen.insert(entry.peer_fp.clone()) {
+            out.push(entry.clone());
+        }
+    }
+    out
+}
+
+/// Clear the entire history and persist the empty file.
+pub fn clear() {
+    let mut guard = store().write().unwrap();
+    guard.clear();
+    save_to_disk(&guard);
+}
+
+/// Find a Missed-candidate entry that matches `call_id` and hasn't been
+/// answered yet. Used by the signal loop to turn "pending incoming" into
+/// "Received" when the user accepts.
+pub fn mark_received_if_pending(call_id: &str) -> bool {
+    let mut guard = store().write().unwrap();
+    for entry in guard.iter_mut().rev() {
+        if entry.call_id == call_id && entry.direction == CallDirection::Missed {
+            entry.direction = CallDirection::Received;
+            save_to_disk(&guard);
+            return true;
+        }
+    }
+    false
+}

desktop/src-tauri/src/main.rs

@@ -0,0 +1,10 @@
+// Desktop binary entry point. All logic lives in `lib.rs` so the same
+// code can be built as a cdylib for Android/iOS via `cargo tauri android build`.
+#![cfg_attr(
+    all(not(debug_assertions), target_os = "windows"),
+    windows_subsystem = "windows"
+)]
+
+fn main() {
+    wzp_desktop_lib::run();
+}

desktop/src-tauri/src/wzp_native.rs

@@ -0,0 +1,148 @@
+//! Runtime binding to the standalone `wzp-native` cdylib.
+//!
+//! See `docs/incident-tauri-android-init-tcb.md` and the top of
+//! `crates/wzp-native/src/lib.rs` for the full story on why this split
+//! exists. Short version: Tauri's desktop cdylib cannot have any C++
+//! compiled into it (via cc::Build) without landing in rust-lang/rust#104707's
+//! staticlib symbol leak, which makes bionic's private `pthread_create`
+//! symbols bind locally and SIGSEGV in `__init_tcb+4` at launch. So all
+//! the Oboe + audio code lives in a standalone `wzp-native` .so built
+//! with `cargo-ndk`, and we dlopen it here at runtime.
+//!
+//! The Library handle lives in a `'static` `OnceLock` for the lifetime of
+//! the process; all function pointers cached below borrow from it safely.
+
+#![cfg(target_os = "android")]
+
+use std::sync::OnceLock;
+
+// ─── Library handle (kept alive forever) ─────────────────────────────────
+
+static LIB: OnceLock<libloading::Library> = OnceLock::new();
+
+// Cached function pointers, resolved once at init(). Each is a raw
+// `extern "C"` fn pointer with effectively `'static` lifetime because
+// LIB is a OnceLock that never drops.
+static VERSION: OnceLock<unsafe extern "C" fn() -> i32> = OnceLock::new();
+static HELLO: OnceLock<unsafe extern "C" fn(*mut u8, usize) -> usize> = OnceLock::new();
+static AUDIO_START: OnceLock<unsafe extern "C" fn() -> i32> = OnceLock::new();
+static AUDIO_START_BT: OnceLock<unsafe extern "C" fn() -> i32> = OnceLock::new();
+static AUDIO_STOP: OnceLock<unsafe extern "C" fn()> = OnceLock::new();
+static AUDIO_READ_CAPTURE: OnceLock<unsafe extern "C" fn(*mut i16, usize) -> usize> = OnceLock::new();
+static AUDIO_WRITE_PLAYOUT: OnceLock<unsafe extern "C" fn(*const i16, usize) -> usize> = OnceLock::new();
+static AUDIO_IS_RUNNING: OnceLock<unsafe extern "C" fn() -> i32> = OnceLock::new();
+static AUDIO_CAPTURE_LATENCY: OnceLock<unsafe extern "C" fn() -> f32> = OnceLock::new();
+static AUDIO_PLAYOUT_LATENCY: OnceLock<unsafe extern "C" fn() -> f32> = OnceLock::new();
+
+/// Load `libwzp_native.so` and resolve every exported function we use.
+/// Call this once at app startup (from the Tauri `setup()` callback).
+/// Subsequent calls are no-ops.
+pub fn init() -> Result<(), String> {
+    if LIB.get().is_some() {
+        return Ok(());
+    }
+
+    // Open the sibling cdylib. The Android dynamic linker searches
+    // /data/app/<pkg>/lib/arm64/ which gradle populates from jniLibs.
+    let lib = unsafe { libloading::Library::new("libwzp_native.so") }
+        .map_err(|e| format!("dlopen libwzp_native.so: {e}"))?;
+
+    // Stash the Library into the OnceLock first so all Symbol lookups
+    // below borrow from the 'static reference rather than a local.
+    LIB.set(lib).map_err(|_| "wzp_native::LIB already set")?;
+    let lib_ref: &'static libloading::Library = LIB.get().unwrap();
+
+    unsafe {
+        macro_rules! resolve {
+            ($cell:expr, $ty:ty, $name:expr) => {{
+                let sym: libloading::Symbol<$ty> = lib_ref.get($name)
+                    .map_err(|e| format!("dlsym {}: {e}", core::str::from_utf8($name).unwrap_or("?")))?;
+                // Dereference the Symbol to extract the raw fn pointer;
+                // it stays valid because lib_ref is 'static.
+                $cell.set(*sym).map_err(|_| format!("{} already set", core::str::from_utf8($name).unwrap_or("?")))?;
+            }};
+        }
+
+        resolve!(VERSION, unsafe extern "C" fn() -> i32, b"wzp_native_version");
+        resolve!(HELLO, unsafe extern "C" fn(*mut u8, usize) -> usize, b"wzp_native_hello");
+        resolve!(AUDIO_START, unsafe extern "C" fn() -> i32, b"wzp_native_audio_start");
+        resolve!(AUDIO_START_BT, unsafe extern "C" fn() -> i32, b"wzp_native_audio_start_bt");
+        resolve!(AUDIO_STOP, unsafe extern "C" fn(), b"wzp_native_audio_stop");
+        resolve!(AUDIO_READ_CAPTURE, unsafe extern "C" fn(*mut i16, usize) -> usize, b"wzp_native_audio_read_capture");
+        resolve!(AUDIO_WRITE_PLAYOUT, unsafe extern "C" fn(*const i16, usize) -> usize, b"wzp_native_audio_write_playout");
+        resolve!(AUDIO_IS_RUNNING, unsafe extern "C" fn() -> i32, b"wzp_native_audio_is_running");
+        resolve!(AUDIO_CAPTURE_LATENCY, unsafe extern "C" fn() -> f32, b"wzp_native_audio_capture_latency_ms");
+        resolve!(AUDIO_PLAYOUT_LATENCY, unsafe extern "C" fn() -> f32, b"wzp_native_audio_playout_latency_ms");
+    }
+
+    Ok(())
+}
+
+/// Is `init()` done and all symbols cached?
+pub fn is_loaded() -> bool {
+    AUDIO_START.get().is_some()
+}
+
+// ─── Smoke-test accessors ────────────────────────────────────────────────
+
+pub fn version() -> i32 {
+    VERSION.get().map(|f| unsafe { f() }).unwrap_or(-1)
+}
+
+pub fn hello() -> String {
+    let Some(f) = HELLO.get() else { return String::new(); };
+    let mut buf = [0u8; 64];
+    let n = unsafe { f(buf.as_mut_ptr(), buf.len()) };
+    String::from_utf8_lossy(&buf[..n]).into_owned()
+}
+
+// ─── Audio accessors ─────────────────────────────────────────────────────
+
+/// Start the Oboe capture + playout streams. Returns `Err(code)` on
+/// failure. Idempotent on the wzp-native side.
+pub fn audio_start() -> Result<(), i32> {
+    let f = AUDIO_START.get().ok_or(-100_i32)?;
+    let ret = unsafe { f() };
+    if ret == 0 { Ok(()) } else { Err(ret) }
+}
+
+/// Start Oboe in Bluetooth SCO mode — capture skips sample rate and
+/// input preset so the system routes to the BT SCO device natively.
+pub fn audio_start_bt() -> Result<(), i32> {
+    let f = AUDIO_START_BT.get().ok_or(-100_i32)?;
+    let ret = unsafe { f() };
+    if ret == 0 { Ok(()) } else { Err(ret) }
+}
+
+/// Stop both streams. Safe to call even if not running.
+pub fn audio_stop() {
+    if let Some(f) = AUDIO_STOP.get() {
+        unsafe { f() };
+    }
+}
+
+/// Read captured i16 PCM into `out`. Returns bytes actually copied.
+pub fn audio_read_capture(out: &mut [i16]) -> usize {
+    let Some(f) = AUDIO_READ_CAPTURE.get() else { return 0; };
+    unsafe { f(out.as_mut_ptr(), out.len()) }
+}
+
+/// Write i16 PCM into the playout ring. Returns samples enqueued.
+pub fn audio_write_playout(input: &[i16]) -> usize {
+    let Some(f) = AUDIO_WRITE_PLAYOUT.get() else { return 0; };
+    unsafe { f(input.as_ptr(), input.len()) }
+}
+
+pub fn audio_is_running() -> bool {
+    AUDIO_IS_RUNNING.get().map(|f| unsafe { f() } != 0).unwrap_or(false)
+}
+
+#[allow(dead_code)]
+pub fn audio_capture_latency_ms() -> f32 {
+    AUDIO_CAPTURE_LATENCY.get().map(|f| unsafe { f() }).unwrap_or(0.0)
+}
+
+#[allow(dead_code)]
+pub fn audio_playout_latency_ms() -> f32 {
+    AUDIO_PLAYOUT_LATENCY.get().map(|f| unsafe { f() }).unwrap_or(0.0)
+}

desktop/src-tauri/tauri.conf.json

@@ -0,0 +1,36 @@
+{
+  "productName": "WarzonePhone",
+  "version": "0.1.0",
+  "identifier": "com.wzp.desktop",
+  "build": {
+    "frontendDist": "../dist",
+    "devUrl": "http://localhost:1420",
+    "beforeDevCommand": "npm run dev",
+    "beforeBuildCommand": "npm run build"
+  },
+  "app": {
+    "windows": [
+      {
+        "title": "WarzonePhone",
+        "width": 400,
+        "height": 640,
+        "resizable": true,
+        "minWidth": 360,
+        "minHeight": 500
+      }
+    ],
+    "security": {
+      "csp": null
+    }
+  },
+  "bundle": {
+    "active": true,
+    "targets": "all",
+    "icon": [
+      "icons/icon.png"
+    ],
+    "android": {
+      "minSdkVersion": 26
+    }
+  }
+}

desktop/src/identicon.ts

@@ -0,0 +1,110 @@
+/**
+ * Deterministic identicon generator — creates a unique symmetric pattern
+ * from a hex fingerprint string, similar to MetaMask's Jazzicon / Ethereum blockies.
+ *
+ * Returns an SVG data URL that can be used as an <img> src.
+ */
+
+function hashBytes(hex: string): number[] {
+  const clean = hex.replace(/[^0-9a-fA-F]/g, "");
+  const bytes: number[] = [];
+  for (let i = 0; i < clean.length; i += 2) {
+    bytes.push(parseInt(clean.substring(i, i + 2), 16));
+  }
+  // Pad to at least 16 bytes
+  while (bytes.length < 16) bytes.push(0);
+  return bytes;
+}
+
+function hslToRgb(h: number, s: number, l: number): [number, number, number] {
+  s /= 100;
+  l /= 100;
+  const k = (n: number) => (n + h / 30) % 12;
+  const a = s * Math.min(l, 1 - l);
+  const f = (n: number) =>
+    l - a * Math.max(-1, Math.min(k(n) - 3, Math.min(9 - k(n), 1)));
+  return [
+    Math.round(f(0) * 255),
+    Math.round(f(8) * 255),
+    Math.round(f(4) * 255),
+  ];
+}
+
+export function generateIdenticon(
+  fingerprint: string,
+  size: number = 36
+): string {
+  const bytes = hashBytes(fingerprint);
+
+  // Derive colors from first bytes
+  const hue1 = (bytes[0] * 360) / 256;
+  const hue2 = ((bytes[1] * 360) / 256 + 120) % 360;
+  const [r1, g1, b1] = hslToRgb(hue1, 65, 35); // dark bg
+  const [r2, g2, b2] = hslToRgb(hue2, 70, 55); // bright fg
+
+  const bg = `rgb(${r1},${g1},${b1})`;
+  const fg = `rgb(${r2},${g2},${b2})`;
+
+  // 5x5 grid, left-right symmetric (only need 3 columns)
+  const grid: boolean[][] = [];
+  for (let y = 0; y < 5; y++) {
+    const row: boolean[] = [];
+    for (let x = 0; x < 3; x++) {
+      const byteIdx = 2 + y * 3 + x;
+      row.push(bytes[byteIdx % bytes.length] > 128);
+    }
+    // Mirror: col 3 = col 1, col 4 = col 0
+    grid.push([row[0], row[1], row[2], row[1], row[0]]);
+  }
+
+  // Render SVG
+  const cellSize = size / 5;
+  const r = size * 0.12; // border radius
+  let rects = "";
+  for (let y = 0; y < 5; y++) {
+    for (let x = 0; x < 5; x++) {
+      if (grid[y][x]) {
+        rects += `<rect x="${x * cellSize}" y="${y * cellSize}" width="${cellSize}" height="${cellSize}" fill="${fg}"/>`;
+      }
+    }
+  }
+
+  const svg = `<svg xmlns="http://www.w3.org/2000/svg" width="${size}" height="${size}" viewBox="0 0 ${size} ${size}">
+    <rect width="${size}" height="${size}" rx="${r}" fill="${bg}"/>
+    ${rects}
+  </svg>`;
+
+  return `data:image/svg+xml,${encodeURIComponent(svg)}`;
+}
+
+/**
+ * Create an <img> element with the identicon.
+ * Click copies the fingerprint to clipboard.
+ */
+export function createIdenticonEl(
+  fingerprint: string,
+  size: number = 36,
+  clickToCopy: boolean = true
+): HTMLImageElement {
+  const img = document.createElement("img");
+  img.src = generateIdenticon(fingerprint, size);
+  img.width = size;
+  img.height = size;
+  img.style.borderRadius = `${size * 0.12}px`;
+  img.style.cursor = clickToCopy ? "pointer" : "default";
+  img.title = fingerprint;
+
+  if (clickToCopy && fingerprint) {
+    img.addEventListener("click", (e) => {
+      e.stopPropagation();
+      navigator.clipboard.writeText(fingerprint).then(() => {
+        img.style.outline = "2px solid #4ade80";
+        setTimeout(() => {
+          img.style.outline = "";
+        }, 600);
+      });
+    });
+  }
+
+  return img;
+}

desktop/tsconfig.json

@@ -0,0 +1,15 @@
+{
+  "compilerOptions": {
+    "target": "ESNext",
+    "module": "ESNext",
+    "moduleResolution": "bundler",
+    "strict": true,
+    "esModuleInterop": true,
+    "skipLibCheck": true,
+    "forceConsistentCasingInFileNames": true,
+    "resolveJsonModule": true,
+    "allowImportingTsExtensions": true,
+    "noEmit": true
+  },
+  "include": ["src"]
+}

desktop/vite.config.ts

@@ -0,0 +1,15 @@
+import { defineConfig } from "vite";
+
+export default defineConfig({
+  clearScreen: false,
+  server: {
+    port: 1420,
+    strictPort: true,
+  },
+  envPrefix: ["VITE_", "TAURI_"],
+  build: {
+    target: "esnext",
+    minify: !process.env.TAURI_DEBUG ? "esbuild" : false,
+    sourcemap: !!process.env.TAURI_DEBUG,
+  },
+});

docs/ADMINISTRATION.md

@@ -625,3 +625,123 @@ curl -s http://relay-host:9090/metrics | grep wzp_relay_active_sessions
 # Check federation probe health
 curl -s http://relay-host:9090/metrics | grep wzp_probe_up
 ```
+
+## Build Pipelines
+
+All production artifacts (Android APK, Linux x86_64 binaries, Windows `.exe`) are built on **SepehrHomeserverdk** using Docker, not on developer workstations. The pipelines are fire-and-forget: a local script invokes a `tmux` session on the remote, the build runs in a Docker container, and the artifact is uploaded to `paste.dk.manko.yoga` (rustypaste) with a notification sent to `ntfy.sh/wzp` on start and completion.
+
+### Docker images
+
+Two long-lived images live on the remote:
+
+| Image | Used by | Base | Key contents |
+|---|---|---|---|
+| `wzp-android-builder` | Android APK (Tauri mobile + legacy Kotlin), Linux x86_64 relay/CLI | Debian bookworm | Rust stable with Android targets, cargo-ndk, NDK 26.1, Android SDK (API 34 + 35 + 36), JDK 17, Gradle 8.5, Node.js 20, cmake, ninja, tauri-cli 2.x |
+| `wzp-windows-builder` | Windows x86_64 `.exe` | Debian bookworm | Rust stable with `x86_64-pc-windows-msvc` target, cargo-xwin (with pre-warmed MSVC CRT + Windows SDK cache), Node.js 20, cmake, ninja, clang, lld, nasm |
+
+Both images are rebuilt rarely — once the base toolchain is stable, rebuilds are only needed to pick up new dependencies or security patches.
+
+**Rebuilding an image** (fire-and-forget, ~10 min on a warm base):
+
+```bash
+# Windows
+./scripts/build-windows-docker.sh --image-build
+
+# Android (upload and rebuild handled by the Android build script itself — see
+# its --image-build flag or equivalent)
+```
+
+The `--image-build` flag uploads the local Dockerfile to the remote, kicks off `docker build` under `nohup`, and returns immediately. Monitor with:
+
+```bash
+ssh SepehrHomeserverdk 'tail -f /tmp/wzp-windows-image-build.log'
+```
+
+### Pipeline: Android APK (Tauri Mobile)
+
+```bash
+./scripts/build-tauri-android.sh                  # Full: pull + build + upload + notify
+./scripts/build-tauri-android.sh --no-pull        # Skip git fetch
+./scripts/build-tauri-android.sh --clean          # Force-clean Rust target
+```
+
+- **Branch**: `android-rewrite`
+- **Image**: `wzp-android-builder`
+- **Build command**: `cargo tauri android build --release`
+- **Output**: `wzp-release.apk` → uploaded to rustypaste
+- **Notifications**: start + completion to `ntfy.sh/wzp`
+- **Remote artifact path**: `/mnt/storage/manBuilder/data/cache-android/target/…/release/app-release.apk`
+
+### Pipeline: Linux x86_64 (relay + CLI + bench + web)
+
+```bash
+./scripts/build-linux-docker.sh                   # Fire-and-forget
+./scripts/build-linux-docker.sh --no-pull         # Skip git fetch
+./scripts/build-linux-docker.sh --clean           # Force-clean target
+./scripts/build-linux-docker.sh --install         # Wait for completion and download locally
+```
+
+- **Branch**: `feat/android-voip-client` (script default — override by editing the script or passing an env var)
+- **Image**: `wzp-android-builder` (shared, not a separate Linux-only image)
+- **Targets built**: `wzp-relay`, `wzp-client`, `wzp-client-audio` (with `--features audio`), `wzp-web`, `wzp-bench`
+- **Output**: `wzp-linux-x86_64.tar.gz` with all five binaries → uploaded to rustypaste
+- **Local landing dir** (with `--install`): `target/linux-x86_64/`
+
+### Pipeline: Windows x86_64 (`wzp-desktop.exe`)
+
+```bash
+./scripts/build-windows-docker.sh                 # Full: pull + build + download locally
+./scripts/build-windows-docker.sh --no-pull       # Skip git fetch
+./scripts/build-windows-docker.sh --rust          # Force-clean target-windows cache
+./scripts/build-windows-docker.sh --image-build   # Rebuild the Docker image (fire-and-forget)
+```
+
+- **Branch**: `feat/desktop-audio-rewrite`
+- **Image**: `wzp-windows-builder`
+- **Build command**: `cargo xwin build --release --target x86_64-pc-windows-msvc --bin wzp-desktop`
+- **Output**: `wzp-desktop.exe` (~16 MB) → downloaded to `target/windows-exe/wzp-desktop.exe`, also uploaded to rustypaste
+- **Target cache volume**: `target-windows` (separate from the Android target cache to avoid triple cross-contamination)
+- **Shared cache volumes**: `cargo-registry`, `cargo-git` (shared with Android — both pipelines pull the same crates)
+
+**A/B-preserving workflow** for testing audio backends: rename the prior `.exe` before re-running the build, so both coexist:
+
+```bash
+# Preserve prior build as the noAEC baseline
+mv target/windows-exe/wzp-desktop.exe target/windows-exe/wzp-desktop-noAEC.exe
+./scripts/build-windows-docker.sh
+ls -la target/windows-exe/
+# wzp-desktop-noAEC.exe  (previous build)
+# wzp-desktop.exe        (new build)
+```
+
+### Alternative pipeline: Windows via Hetzner Cloud VPS
+
+For situations where Docker image rebuilds would be disruptive, or for one-shot debug builds on a clean machine:
+
+```bash
+./scripts/build-windows-cloud.sh                  # Full: create VM → build → download → destroy
+./scripts/build-windows-cloud.sh --prepare        # Create VM + install deps, don't build
+./scripts/build-windows-cloud.sh --build          # Build on existing VM
+./scripts/build-windows-cloud.sh --transfer       # Download .exe from existing VM
+./scripts/build-windows-cloud.sh --destroy        # Delete the VM
+WZP_KEEP_VM=1 ./scripts/build-windows-cloud.sh    # Don't auto-destroy after successful build
+```
+
+- **Provider**: Hetzner Cloud
+- **Default server type**: `cx33` (8 GB RAM, 8 vCPU — `cx23` with 4 GB OOMs on the tauri+rustls cross-compile)
+- **Image**: `ubuntu-24.04`
+- **SSH key**: must be named `wz` in Hetzner and loaded in the local ssh-agent
+- **Reminder**: set `WZP_KEEP_VM=1` for multi-build sessions, then **remember to `--destroy` at end of day** so the VM isn't left running overnight. This is tracked in the auto-memory as `feedback_keep_windows_builder_vm.md`.
+
+### Notifications
+
+All pipelines post to `https://ntfy.sh/wzp`. Subscribe from your phone via the [ntfy.sh app](https://ntfy.sh/) to get push notifications on build start/success/failure. Messages include the short git hash and the rustypaste URL on success:
+
+```
+WZP Windows build OK [03a80a3] (16M)
+https://paste.dk.manko.yoga/<uuid>/wzp-desktop.exe
+```
+
+### Rustypaste credentials
+
+Build pipelines read `rusty_address` and `rusty_auth_token` from the `.env` file at `/mnt/storage/manBuilder/.env` on SepehrHomeserverdk. Local scripts that upload directly (`build-windows-cloud.sh` when run in `--transfer` mode) read from `~/.wzp/rustypaste.env` with the same variable names. Both files must be kept in sync manually if rotated.

docs/ARCHITECTURE.md

@@ -103,11 +103,13 @@ sequenceDiagram
     participant RNN as RNNoise<br/>(2 x 480)
     participant VAD as SilenceDetector
     participant Codec as Opus / Codec2
+    participant DT as DredTuner<br/>(wzp-proto)
     participant FEC as RaptorQ FEC
     participant INT as Interleaver<br/>(depth=3)
     participant HDR as MediaHeader<br/>(12B or Mini 4B)
     participant Enc as ChaCha20-Poly1305
     participant QUIC as QUIC Datagram
+    participant QPS as QuinnPathSnapshot
 
     Mic->>Ring: f32 x 512 (macOS callback)
     Ring->>Ring: Accumulate to 960 samples
@@ -118,10 +120,19 @@ sequenceDiagram
     else Silence (>100ms)
         VAD->>Codec: ComfortNoise (every 200ms)
     end
+
+    Note over QPS,DT: Every 25 frames (~500ms)
+    QPS->>DT: loss_pct, rtt_ms, jitter_ms
+    DT->>Codec: set_dred_duration() + set_expected_loss()
+
+    alt Opus tier (any bitrate)
+        Codec->>HDR: Compressed bytes + DRED side-channel (no RaptorQ)
+    else Codec2 tier
         Codec->>FEC: Compressed bytes (pad to 256B symbol)
         FEC->>FEC: Accumulate block (5-10 symbols)
         FEC->>INT: Source + repair symbols
         INT->>HDR: Interleaved packets
+    end
     HDR->>Enc: Header as AAD
     Enc->>QUIC: Encrypted payload + 16B tag
 ```
@@ -134,6 +145,9 @@ sequenceDiagram
 - Silence detection uses VAD + 100ms hangover before switching to ComfortNoise
 - FEC symbols are padded to **256 bytes** with a 2-byte LE length prefix
 - MiniHeaders (4 bytes) replace full headers (12 bytes) for 49 of every 50 frames
+- DRED tuner polls quinn path stats every 25 frames (~500ms) and adjusts DRED lookback duration continuously
+- Opus tiers bypass RaptorQ entirely -- DRED handles loss recovery at the codec layer
+- Opus6k DRED window: 1040ms (maximum libopus allows)
 
 ## Audio Decode Pipeline
 
@@ -154,13 +168,30 @@ sequenceDiagram
     Dec->>AR: Decrypt (header = AAD)
     AR->>AR: Check seq window (reject replay)
     AR->>HDR: Verified packet
+
+    alt Opus packet
+        HDR->>JIT: Direct to jitter buffer (no FEC/interleave)
+    else Codec2 packet
         HDR->>DEINT: MediaHeader + payload
         DEINT->>FEC: Reordered symbols by block
         FEC->>FEC: Attempt decode (need K of K+R)
         FEC->>JIT: Recovered audio frames
+    end
+
     JIT->>JIT: BTreeMap ordered by seq
     JIT->>JIT: Wait until depth >= target
+
+    alt Packet present
         JIT->>Codec: Pop lowest seq frame
+    else Packet missing (Opus)
+        JIT->>Codec: DRED reconstruction (neural)
+        alt DRED fails or unavailable
+            Codec->>Codec: Classical PLC fallback
+        end
+    else Packet missing (Codec2)
+        Codec->>Codec: Classical PLC
+    end
+
     Codec->>Ring: PCM i16 x 960
     Ring->>SPK: Audio callback pulls samples
 ```
@@ -172,6 +203,8 @@ sequenceDiagram
 - Jitter buffer target: **10 packets (200ms)** for client, **50 packets (1s)** for relay
 - Desktop client uses **direct playout** (no jitter buffer) with lock-free ring
 - Codec2 frames at 8 kHz are resampled to 48 kHz transparently
+- DRED reconstruction: on packet loss, decoder tries neural DRED reconstruction before falling back to classical PLC
+- Jitter-spike detection pre-emptively boosts DRED to ceiling when jitter variance spikes >30%
 
 ## Relay SFU Forwarding
 
@@ -211,6 +244,7 @@ graph TB
 3. If one send fails, the relay continues to the next participant (best-effort)
 4. The relay never decodes or re-encodes audio (preserves E2E encryption)
 5. With trunking enabled, packets to the same receiver are batched into TrunkFrames (flushed every 5ms)
+6. Relay tracks per-participant quality from QualityReport trailers and broadcasts `QualityDirective` when the room-wide tier degrades (coordinated codec switching)
 
 ## Federation Topology
 
@@ -348,7 +382,7 @@ Used for 49 of every 50 frames (~1s cycle). Saves 8 bytes per packet (67% header
   [session_id: 2][len: u16][payload: len]  x count
 ```
 
-Packs multiple session packets into one QUIC datagram. Maximum 10 entries or 1200 bytes, flushed every 5ms.
+Packs multiple session packets into one QUIC datagram. Maximum 10 entries or PMTUD-discovered MTU (starts at 1200, grows to ~1452 on Ethernet), flushed every 5ms.
 
 ### QualityReport (4 bytes, optional trailer)
 
@@ -361,6 +395,40 @@ Byte 3: bitrate_cap_kbps (0-255 kbps)
 
 Appended to a media packet when the Q flag is set in the MediaHeader.
 
+## Path MTU Discovery
+
+Quinn's PLPMTUD is enabled with:
+- `initial_mtu`: 1200 bytes (QUIC minimum, always safe)
+- `upper_bound`: 1452 bytes (Ethernet minus IP/UDP/QUIC headers)
+- `interval`: 300s (re-probe every 5 minutes)
+- `black_hole_cooldown`: 30s (faster retry on lossy links)
+
+The discovered MTU is exposed via `QuinnPathSnapshot::current_mtu` and used by:
+- `TrunkedForwarder`: refreshes `max_bytes` on every send to fill larger datagrams
+- Future video framer: larger MTU = fewer application-layer fragments per frame
+
+## Continuous DRED Tuning
+
+Instead of locking DRED duration to 3 discrete quality tiers, the `DredTuner` (in `wzp-proto::dred_tuner`) maps live path quality to a continuous DRED duration:
+
+| Input | Source | Update Rate |
+|-------|--------|-------------|
+| Loss % | `QuinnPathSnapshot::loss_pct` (from quinn ACK frames) | Every 25 packets (~500ms) |
+| RTT ms | `QuinnPathSnapshot::rtt_ms` (quinn congestion controller) | Every 25 packets |
+| Jitter ms | `PathMonitor::jitter_ms` (EWMA of RTT variance) | Every 25 packets |
+
+### Mapping Logic
+
+- **Baseline**: codec-tier default (Studio=100ms, Good=200ms, Degraded=500ms)
+- **Ceiling**: codec-tier max (Studio=300ms, Good=500ms, Degraded=1040ms)
+- **Continuous**: linear interpolation between baseline and ceiling based on loss (0%->baseline, 40%->ceiling)
+- **RTT phantom loss**: high RTT (>200ms) adds phantom loss contribution to keep DRED generous
+- **Jitter spike**: >30% EWMA spike pre-emptively boosts to ceiling for ~5s cooldown
+
+### Output
+
+`DredTuning { dred_frames: u8, expected_loss_pct: u8 }` -> fed to `CallEncoder::apply_dred_tuning()` -> `OpusEncoder::set_dred_duration()` + `set_expected_loss()`
+
 ## Signal Message Handshake Flow
 
 ```mermaid
@@ -872,3 +940,135 @@ warzonePhone/
 | wzp-relay | 40 + 4 integration | Room ACL, session mgmt, metrics, probes, mesh, trunking |
 | wzp-client | 30 + 2 integration | Encoder/decoder, quality adapter, silence, drift, sweep |
 | wzp-web | 2 | Metrics |
+
+## Audio Backend Architecture (Platform Matrix)
+
+WarzonePhone's audio I/O goes through one of four backends depending on the target platform and feature flags. All backends expose the same public API (`AudioCapture::start() → AudioCapture { ring(), stop() }`) via conditional re-exports in `crates/wzp-client/src/lib.rs`, so the `CallEngine` above the audio layer doesn't know or care which backend is running.
+
+```
+            ┌─────────────────────────────────────────────┐
+            │         CallEngine (platform-agnostic)       │
+            │    reads PCM from AudioCapture::ring()       │
+            │    writes PCM to   AudioPlayback::ring()     │
+            └────────────────────┬────────────────────────┘
+                                 │
+           ┌─────────────────────┼─────────────────────┐
+           │                     │                     │
+           ▼                     ▼                     ▼
+   ┌───────────────┐    ┌────────────────┐    ┌───────────────┐
+   │   audio_io    │    │  audio_vpio    │    │ audio_wasapi  │
+   │   (CPAL)      │    │ (Core Audio    │    │   (Windows    │
+   │               │    │  VoiceProc IO) │    │  IAudioClient2│
+   │ All platforms │    │   macOS only   │    │   Windows     │
+   │  (baseline)   │    │   feature=vpio │    │ feature=      │
+   │               │    │                │    │  windows-aec  │
+   └───────────────┘    └────────────────┘    └───────────────┘
+                                                       │
+                                                       ▼ on Android only
+                                               ┌───────────────┐
+                                               │  wzp-native   │
+                                               │ (Oboe bridge  │
+                                               │  via dlopen)  │
+                                               │               │
+                                               │ Android only  │
+                                               │  libloading   │
+                                               └───────────────┘
+```
+
+### Backend selection matrix
+
+| Platform | Capture | Playback | OS AEC | Feature flags |
+|---|---|---|---|---|
+| macOS | VoiceProcessingIO (native Core Audio) | CPAL | **Yes** — Apple's hardware-accelerated AEC (same AEC as FaceTime, iMessage audio, Voice Memos) | `audio`, `vpio` |
+| Windows (AEC build) | Direct WASAPI with `AudioCategory_Communications` | CPAL | **Yes** — Windows routes the capture stream through the driver's communications APO chain (AEC + NS + AGC), driver-dependent quality | `audio`, `windows-aec` |
+| Windows (baseline) | CPAL (WASAPI shared mode) | CPAL | No | `audio` |
+| Linux | CPAL (ALSA / PulseAudio) | CPAL | No | `audio` |
+| Android (Tauri Mobile) | Oboe via `wzp-native` cdylib, `Usage::VoiceCommunication` + `MODE_IN_COMMUNICATION` | Same Oboe stream | Depends on device (some Android devices apply AEC to the voice-communication stream, most do not) | none (`wzp-client` compiled with `default-features = false`) |
+
+### Why `wzp-native` is a standalone cdylib
+
+On Android, the audio backend lives in a separate cdylib crate (`crates/wzp-native`) that `wzp-desktop`'s lib crate loads at runtime via `libloading`. It is **not** linked as a regular Rust dep.
+
+This is deliberate. rust-lang/rust#104707 documents that a crate with `crate-type = ["cdylib", "staticlib"]` leaks non-exported symbols from the staticlib into the cdylib. On Android, that caused Bionic's private `__init_tcb` / `pthread_create` symbols to be bound LOCALLY inside our `.so` instead of resolved dynamically against `libc.so` at `dlopen` time — which crashed the app at launch as soon as `tao` tried to `std::thread::spawn()` from the JNI `onCreate` callback.
+
+Keeping `wzp-native` in its own cdylib and loading it via `libloading` means:
+
+1. The app's own `.so` has `crate-type = ["cdylib", "rlib"]` only — no `staticlib`, no symbol leak.
+2. `libwzp_native.so` is loaded via `System.loadLibrary` from the JVM side (or `dlopen` from Rust), which triggers the normal Bionic resolver and binds all private symbols against `libc.so` at load time.
+3. The C/C++ Oboe bridge is fully isolated inside `libwzp_native.so`'s symbol space — no chance of its archives leaking into `wzp-desktop`'s `.so`.
+
+See `docs/BRANCH-android-rewrite.md` for the full incident postmortem and `docs/incident-tauri-android-init-tcb.md` for the debug log.
+
+### Vendored `audiopus_sys` for libopus / clang-cl cross-compile
+
+The workspace root carries a vendored copy of `audiopus_sys` at `vendor/audiopus_sys/` with a patched `opus/CMakeLists.txt`. This is needed because libopus 1.3.1 gates its per-file `-msse4.1` / `-mssse3` `COMPILE_FLAGS` behind `if(NOT MSVC)`, and under `clang-cl` (used by `cargo-xwin` for Windows cross-compiles) CMake sets `MSVC=1` unconditionally — so the SIMD source files compile without the required target feature and fail to link the intrinsic `always_inline` functions.
+
+The patch introduces an `MSVC_CL` variable that is true only for real `cl.exe` (distinguished via `CMAKE_C_COMPILER_ID STREQUAL "MSVC"`), and flips the eight `if(NOT MSVC)` SIMD guards to `if(NOT MSVC_CL)` so clang-cl gets the GCC-style per-file flags. Wired in via `[patch.crates-io] audiopus_sys = { path = "vendor/audiopus_sys" }` at the workspace root.
+
+This does not affect macOS or Linux builds — on those platforms `MSVC=0` everywhere so the patched logic behaves identically to upstream.
+
+Upstream tracking: xiph/opus#256, xiph/opus PR #257 (both stale).
+
+## Network Awareness (Android)
+
+The adaptive quality controller (`AdaptiveQualityController` in `wzp-proto`) supports proactive network-aware adaptation via `signal_network_change(NetworkContext)`. On Android, this is fed by `NetworkMonitor.kt` which wraps `ConnectivityManager.NetworkCallback`.
+
+```
+ConnectivityManager
+       │ onCapabilitiesChanged / onLost
+       ▼
+NetworkMonitor.kt  ──classify──►  type: Int (WiFi=0, LTE=1, 5G=2, 3G=3)
+       │ onNetworkChanged(type, bw)
+       ▼
+CallViewModel  ──►  WzpEngine.onNetworkChanged()
+                        │ JNI
+                        ▼
+                    jni_bridge.rs
+                        │
+                        ▼
+                    EngineState.pending_network_type  (AtomicU8, lock-free)
+                        │ polled every ~20ms
+                        ▼
+                    recv task: quality_ctrl.signal_network_change(ctx)
+                        │
+                        ├─ WiFi → Cellular: preemptive 1-tier downgrade
+                        ├─ Any change: 10s FEC boost (+0.2 ratio)
+                        └─ Cellular: faster downgrade thresholds (2 vs 3)
+```
+
+Cellular generation is approximated from `getLinkDownstreamBandwidthKbps()` to avoid requiring `READ_PHONE_STATE` permission.
+
+## Audio Routing (Android)
+
+Both Android app variants support 3-way audio routing: **Earpiece → Speaker → Bluetooth SCO**.
+
+### Audio Mode Lifecycle
+
+`MODE_IN_COMMUNICATION` is set by the Rust call engine (via JNI `AudioManager.setMode()`) right before Oboe streams open — NOT at app launch. Restored to `MODE_NORMAL` when the call ends. This prevents hijacking system audio routing (music, BT A2DP) before a call is active.
+
+### Native Kotlin App
+
+`AudioRouteManager.kt` handles device detection (via `AudioDeviceCallback`), SCO lifecycle, and auto-fallback on BT disconnect. `CallViewModel.cycleAudioRoute()` cycles through available routes.
+
+### Tauri Desktop App
+
+`android_audio.rs` provides JNI bridges to `AudioManager` for speakerphone and Bluetooth SCO control. After each route change, Oboe streams are stopped and restarted via `spawn_blocking`.
+
+```
+User tap ──► cycleAudioRoute()
+                │
+                ├─ Earpiece: setSpeakerphoneOn(false) + clearCommunicationDevice()
+                ├─ Speaker:  setSpeakerphoneOn(true)
+                └─ BT SCO:   setCommunicationDevice(bt_device)  [API 31+]
+                │              fallback: startBluetoothSco()     [API < 31]
+                ▼
+            Oboe stop + start_bt() for BT / start() for others
+```
+
+### BT SCO and Oboe
+
+BT SCO only supports 8/16kHz. When `bt_active=1`, Oboe capture skips `setSampleRate(48000)` and `setInputPreset(VoiceCommunication)`, letting the system choose the native BT rate. Oboe's `SampleRateConversionQuality::Best` bridges to our 48kHz ring buffers. Playout uses `Usage::Media` in BT mode to avoid conflicts with the communication device routing.
+
+### Hangup Signal Fix
+
+`SignalMessage::Hangup` now carries an optional `call_id` field. The relay uses it to end only the specific call instead of broadcasting to all active calls for the user — preventing a race where a hangup for call 1 kills a newly-placed call 2.