T2.3-T2.6: BWE guard, relay conformance Tier A/B/C, Prometheus metrics

docs/PRD/reports/T2.6-report.md

@@ -0,0 +1,77 @@
+# T2.6 — Prometheus metrics for conformance
+
+**Status:** Pending Review
+**Agent:** Kimi Code CLI
+**Started:** 2026-05-11T17:45Z
+**Completed:** 2026-05-11T17:55Z
+**Commit:** 846c98e
+**PRD:** ../PRD-relay-conformance.md
+
+## What I changed
+
+- `crates/wzp-relay/src/metrics.rs`:
+  - Updated `conformance_violations: IntCounterVec` labels from `["violation_type"]` to `["tier", "codec_id", "media_type", "verdict"]`.
+  - Added `conformance_bytes: HistogramVec` — packet size distribution, label `media_type`.
+  - Added `conformance_iat_ms: HistogramVec` — inter-arrival time distribution, label `media_type`.
+  - Added `record_conformance(header, payload_len, iat_ms, violation)` helper:
+    - Records bytes + IAT histograms on **every** packet.
+    - Increments violation counter (with full labels) only on violations.
+
+- `crates/wzp-relay/src/room.rs`:
+  - Both `run_participant_plain` and `run_participant_trunked` call `metrics.record_conformance()` on every incoming packet.
+  - `recv_gap_ms` (already computed for gap logging) is reused as the IAT measurement.
+
+## Why these choices
+
+Histograms are recorded per-packet so operators can see the full distribution of traffic, not just the abusive tail. The `media_type` label separates audio, video, data, and control traffic without over-labeling (codec_id on histograms would create too many time-series).
+
+The violation counter uses four labels:
+- `tier` — "A", "B", or "C" (which conformance check failed)
+- `codec_id` — `Debug` representation (e.g., "Opus24k")
+- `media_type` — `Debug` representation (e.g., "Audio")
+- `verdict` — `Debug` representation of `Violation` enum
+
+This gives operators enough dimensions to correlate violations with specific codecs and traffic types.
+
+## Deviations from the task spec
+
+None.
+
+## Verification output
+
+```bash
+$ cargo test -p wzp-relay conformance
+running 10 tests
+...(all 10 pass)...
+
+test result: ok. 10 passed; 0 failed; 0 ignored; 0 measured; 76 filtered out; finished in 0.00s
+```
+
+```bash
+$ cargo test -p wzp-relay
+running 86 tests
+...(all 86 pass)...
+
+test result: ok. 86 passed; 0 failed; 0 ignored; 0 measured; 0 filtered out; finished in 0.01s
+```
+
+## Test summary
+
+- Tests added: 0 (metrics are exercised indirectly by conformance tests)
+- Tests modified: 0
+- `wzp-relay` test count: 86 (unchanged)
+- `cargo clippy -p wzp-relay --lib`: pass (no new warnings)
+- `cargo fmt --all -- --check`: pass
+
+## Risks / follow-ups
+
+- Histogram cardinality is bounded: `media_type` has 4 values, so `conformance_bytes` and `conformance_iat_ms` each produce 4 time-series. Safe for Prometheus.
+- Violation counter cardinality: `tier` (3) × `codec_id` (~9) × `media_type` (4) × `verdict` (3) = ~324 max combinations. In practice, most participants use only 1-2 codecs, so actual cardinality is much lower.
+
+## Reviewer checklist (filled in by reviewer)
+
+- [ ] Code matches PRD intent
+- [ ] Verification output is real
+- [ ] No backward-incompat surprises
+- [ ] Tests cover the new behavior
+- [ ] Approved