Local Whisper STT provider with browser audio normalization and session-safe submission #21

Outcome

Select local-whisper independently of TTS and transcribe browser utterances locally without keys or 9Router. Native agent model routing remains unchanged.

Implementation

Add local-whisper to shared contracts, settings and speech dispatch. Use whisper.cpp /inference (not the OpenAI /v1/audio/transcriptions path). Convert supported browser WebM/MP4 and WAV audio to engine-compatible WAV locally with bounded, cancelable processing.

Acceptance criteria

• Local STT selection/endpoint persist and require no API key or 9Router.
• Actual browser recording formats and WAV normalize correctly and yield real transcripts.
• Local endpoints are loopback-only; redirects and cloud fallback are prohibited.
• Switching STT does not change TTS, native session ID, agent model or permissions; selection is snapshotted per utterance.
• Success submits exactly once; silence, canceled/stale inference and failure submit nothing. Retry never duplicates native turns.
• Timeout, missing engine/model, malformed output and conversion failure have actionable recoverable errors.
• No audio/transcript/credential leakage into logs or retained conversion files.
• Existing cloud speech and native-session contracts remain green; relevant browser E2E passes.

Required TDD workflow

For each behavior: write and run a focused failing test; confirm the failure is for missing behavior; implement minimally; refactor while green. Record actual red/green commands and outcomes. Use deterministic fixture transports/processes for contracts plus a real pinned engine/model smoke test. No personal sessions or recordings as test fixtures.

Completion evidence

Record commit/PR, test commands, engine/model versions and checksums, Mac hardware, real-engine results, and any unverified manual-device gates. Do not mark unverified acceptance criteria complete.

Implementation evidence

Implemented in #25. Automated acceptance, real offline engine results, and reproducible corpus are recorded in docs/local-whisper-qa.md and the PR. Native physical-device acceptance remains in the separate QA issue.