[witbrock]

Merge decision: ready — recorded dictation now supplies conversation vocabulary, preserves drafts through retry/cancellation, and avoids mobile recognition accumulation and playback workarounds.

User outcome

Repairs the repeated progressively longer prefixes and replayed previous dictation observed in the DGX “Von Task Smoke Test”. A microphone recording produces an editable transcript; Send finishes transcription first, failures retain retryable audio, and late results cannot alter another conversation.

Material changes

• Add authenticated MP4/WebM transcription using an enabled OpenAI audio model and bounded visible conversation/draft/concept-name context.
• Add shared composer capture/transcribe/retry/cancel states; use browser recognition explicitly and handle its results as revised snapshots.
• Exclude supported transcription models from chat fallback/selection candidates.
• Restrict Chromium TTS pause/resume to desktop and keep mobile playback within the tap gesture.
• Document the runtime contract; current streaming/turn-taking plan is in JVNAUTOSCI-812, bounded delivery in JVNAUTOSCI-888.

Evidence

• 13 speech/frontend tests passed; neighbouring chat/presenter checks passed (30 tests).
• 13 targeted backend tests passed with VON_USE_MOCK_DB=1: actual upload handler, normal server session binding, model denial, MP4/WebM/context/language transport, safe errors, and three chat-model candidate boundaries.
• A broader backend run reached 36 passes before being interrupted on unrelated database I/O; it is not claimed complete.
• Frontend static lint and diff checks passed.
• Chrome desktop and 390 × 844 viewport exercised the actual composer/template and controller with simulated microphone/provider: failed upload → edit draft → retry preserved the edit and inserted one transcript; cancellation preserved the draft; no mobile horizontal overflow.

Ship boundary

Minimum ship criteria are met for this bounded transport/lifecycle repair. Lost or duplicated draft, stuck microphone, cross-actor context access, or audio-only models entering chat fallback would block delivery. Physical iOS/Android, acoustic accuracy and runtime deployment are separate receipts, not established by browser emulation. Raw DGX microphone audio and recognition events were not retained; the stored transcript and code establish the accumulation failure class.

Streaming hands-free conversation, server TTS, wake words, paired devices and persistent raw-audio retention are outside this PR. Existing primary models and credentials are preserved; a transcription model must be added as an additional authorised model before runtime use.