Perception card audio + synthesis bench validation #33

Phase A is merged on dotty-stackchan main (63d0cc2 + follow-ups 1fe4b52, 2464ff9, 63359e3) and deployed to the RPi 2026-04-28. Bench-verify on hardware (not yet run):

• (a) curl -F file=@clip.wav -F question="Describe what you hear." -H "device-id: <mac>" http://<ZEROCLAW_HOST>:8080/api/audio/explain returns a sensible description and the dashboard Audio section updates within 2 s.
• (b) The chosen AUDIO_CAPTION_MODEL (default google/gemini-2.5-flash) actually accepts the OpenAI-style input_audio content-block on OpenRouter — first call will reveal a 4xx if the model id needs to change, override via env if so.
• (c) Cover camera 65 s, confirm the Perception card's image-age badge flips to text-warning.
• (d) Wait 5 min with a face present, confirm tail /root/zeroclaw-bridge/logs/scene-synthesis-$(date +%Y-%m-%d).ndjson shows a line and the dashboard Environment section appears.
• (e) Walk in front of camera, confirm face_recognized triggers a synthesis emit within ~120 s (min-interval guard).
• (f) find /root/zeroclaw-bridge -name "*.jpg" -o -name "*.wav" -o -name "*.mp3" returns nothing (no media on disk).
• (g) Confirm the robot-modal scene-context panel no longer duplicates the photo / Sees / Hears (per the dedupe in 63359e3) — only the recent-events ring + security cycles should show.

Migrated from local tasks.md on 2026-05-17.