Local speech QA: real-engine offline, browser formats and performance acceptance #23

Outcome

Prove local STT separately from cloud/native-agent operation and publish reproducible evidence.

Acceptance criteria

• Shared synthetic fixtures cover English, technical phrases, noise, silence and maximum supported input.
• Real pinned engine/model transcribes WAV and browser formats with external networking disabled after setup.
• Record checksum/hardware/cold-warm latency/memory and evaluate proposed <=10% clean-English WER and <=3s p95 warm completion for 10s clips.
• Deterministic tests prove cancellation/session changes cannot dispatch stale or duplicate turns.
• Typecheck, build, unit/integration suites and relevant Playwright tests pass.
• Real microphone/browser acceptance is recorded or explicitly outstanding; mocked agent tests are not described as a fully offline native-agent conversation.
• Documentation distinguishes local transcription from independently selected TTS and Codex/Claude model routing.

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.

Verified evidence and remaining work

Implementation PR #25. Real Whisper base.en on Apple M3/24 GiB passed external-network-denied adapter checks. Twenty synthetic 10-second clean-English clips measured 0.60% normalized WER and 0.138s raw-engine warm p95; these are synthetic-voice results, not representative human accuracy. Model SHA/version, scripts, transcripts, timings and limits are committed in docs/local-whisper-qa.md and .json. Both independently managed engines ran together; stop isolation and offline restart passed.

Physical microphone/browser-device or native hotkey/focus/paste and sleep/wake checks remain unperformed. HTTP cancellation prevents output but may leave the engine computing; no immediate compute-interruption claim is made. Noisy silence can hallucinate despite digital-silence guards. This QA issue remains open for those release acceptance gates.