Final measured run: t_perceived 3.75s PASS, full stage table + findings

docs/LATENCY.md

@@ -2,19 +2,20 @@
 
 ## MEASURED — MVP slice, 2026-10-04, cpu-upgrade (8 vCPU cgroup quota, CPU-only)
 
-End-to-end run `ml-intern-freestream-e2e-wincondition` (job `6ac2dfb8404719ba3765168f`):
+End-to-end run `ml-intern-freestream-e2e-wincondition` (final: job `6ac2e0f3404719ba376516e5`):
 livekit-server v1.13.7 dev mode in-container, bot joined via the real CLI,
 test client asked a spoken question, timings measured at the receiving client.
 
 | Metric | Measured | Target | Verdict |
 |---|---|---|---|
-| **t_perceived** (speech-end → first bot audio at client) | **3.73s** | ≤ 4.0s (win condition) | **PASS** |
-| t_ready (CLI join → bot ready, incl. model warmup) | 10.6s | — | — |
-| ASR finalize (whisper tiny int8, 2 threads) | 1.93s | 400–700ms | over budget |
+| **t_perceived** (speech-end → first bot audio at client) | **3.75s** | ≤ 4.0s (win condition) | **PASS** |
+| t_ready (CLI join → bot ready, incl. model + disclosure warmup) | 14.6s | — | — |
+| ASR finalize (whisper tiny int8, 2 threads) | 2.04s | 400–700ms | over budget |
 | TTS first-audio path (cached consent disclosure) | ~0.2s (pre-roll + transport) | — | PASS |
-| Piper in-context RTF (later sentences) | ~1.2–1.9 | <1.0 | over budget |
-| Piper standalone RTF (no co-tenants) | 0.33–0.65 | <1.0 | PASS |
-| Mouth movement during speech | 230 video frames, median luma diff 0.66 | >0 | PASS |
+| Full turn (speech-end → answer finished) | 18.9s (14.9s of audio) | — | pipelined |
+| Piper in-context RTF (later sentences) | ~0.8–1.2 | <1.0 | borderline |
+| Piper standalone RTF (no co-tenants) | 0.55–0.87 | <1.0 | PASS |
+| Mouth movement during speech | 234 video frames, median luma diff 0.70 | >0 | PASS |
 | Replay logs | both hash chains verify | valid | PASS |
 
 Findings from the measurement campaign:
@@ -26,13 +27,15 @@ Findings from the measurement campaign:
 2. **The fixed consent disclosure is pre-synthesizable** (CONSENT.md #2): it
    is static text, so caching it at session start removes Piper entirely from
    the first-audio critical path (mitigation #4 extended).
-3. **Piper runs 2–3× slower in the bot process than standalone** (RTF 1.2–1.9
-   vs 0.33–0.65) — suspected contention with the co-resident whisper model
-   and render loop. Sentence-streaming hides most of it behind playback;
-   fixing the root cause is the next performance item.
+3. **Piper runs ~2× slower in the bot process than standalone** (RTF ~0.8–1.2
+   vs 0.55–0.87) — suspected contention with the co-resident whisper model
+   and render loop. Pipelined sentence synthesis (rest-of-reply synthesizes
+   while sentence 1 plays, mitigation #2) keeps the turn flowing: 14.9s of
+   audio completed in 18.9s wall clock. Fixing the contention is the next
+   performance item.
 4. **whisper tiny vs base accuracy**: base transcribed the bot name
-   "Streamy" correctly; tiny heard "streaming". Latency profile uses tiny;
-   base (+~2s) is the accuracy upgrade path.
+   "Streamy" correctly; tiny heard "Streami"/"streaming". The latency profile
+   uses tiny; base (+~2s ASR) is the accuracy upgrade path.
 
 The tables below are the original design targets, kept for reference.