feat: add reliable local voice dictation

[yaacovcorcos]

Summary

• add private, offline voice dictation to the Scient composer using a packaged whisper.cpp runtime
• keep recording, model management, transcription, IPC, and UI responsibilities separated behind Scient-owned seams
• preserve Claude's original prototype commit, followed by the reliability hardening and final targeted fixes
• provide record, cancel, transcribe-and-insert, and transcribe-and-send interactions with a responsive waveform and concise tooltips

Why

Scientists should be able to dictate prompts without sending microphone audio or credentials to a cloud transcription service. The implementation also needs to remain narrowly integrated with the T3-derived composer so future upstream updates stay manageable.

Reliability and safety

• guards recorder lifecycle ownership so stale cleanup cannot stop a newer recording
• flushes final audio, aborts stale work, and waits deterministically for the native helper to exit
• stages a pinned and checksummed whisper.cpp runtime without committing generated binaries or models
• gives the child process an explicit environment allowlist
• makes model setup cancellable and resumable
• replaces model receipts atomically and uses bounded user-facing errors
• keeps live transcription out of this first reliable release by explicit product decision

User impact

Users can record speech locally, cancel safely, and either insert the transcription into the composer or send it directly. The composer remains editable while recording. The first model setup can be cancelled and resumed later.

Validation

• manual macOS microphone and visual review approved
• formatting and focused lint passed
• typechecks passed for scient-voice, contracts, web, and desktop
• 2,615 affected-package tests passed
• web production build passed
• git diff --check passed

Limitations

• this first release transcribes after recording stops; it does not stream live partial transcription
• transcription quality follows the bundled small multilingual model and can be revisited independently

[github-actions]

Thread transfer impact

✅ Thread transfer remains within every enforced ceiling.

Provider	Metric	Main baseline	This PR	Impact	PR ceiling
Codex	Total thread wire	11.3 KiB	11.3 KiB	−17 B (−0.1%)	15.1 KiB	✅
Codex	Thread snapshot wire	5.5 KiB	5.5 KiB	+3 B (+0.1%)	7.3 KiB	✅
Codex	Live turn WebSocket wire	5.9 KiB	5.9 KiB	−20 B (−0.3%)	7.8 KiB	✅
Codex	Live turn WebSocket decoded	49.7 KiB	49.7 KiB	0 B (0.0%)	66.4 KiB	✅
Codex	Live turn messages	16	16	0 (0.0%)	21	✅
Claude	Total thread wire	11.3 KiB	11.3 KiB	+5 B (+0.0%)	15.1 KiB	✅
Claude	Thread snapshot wire	5.5 KiB	5.5 KiB	−3 B (−0.1%)	7.3 KiB	✅
Claude	Live turn WebSocket wire	5.8 KiB	5.9 KiB	+8 B (+0.1%)	7.8 KiB	✅
Claude	Live turn WebSocket decoded	50.6 KiB	50.6 KiB	0 B (0.0%)	66.4 KiB	✅
Claude	Live turn messages	16	16	0 (0.0%)	21	✅

Baseline: 96de3c3 · PR result: 324b183 · Source CI: success

Scenario and decoded snapshot size

10 historical turns, 5 command tools per turn, 878.9 KiB retained MCP result per historical turn, and a 1.05 MiB retained result in the measured turn.

• Codex decoded thread snapshot: 94.6 KiB
• Claude decoded thread snapshot: 95.4 KiB

Updated in place by a trusted workflow. PR artifacts are strictly validated and never executed.