fix(mobile): preserve dictation through interruptions and failures

[pseudoseed]

Long dictations could silently stop capturing audio, lose their recording after a transcription error, or lose words during cleanup. This change detects native microphone pauses, transcribes interrupted recordings, retries saved audio, and preserves the original file across recorder resets. It fixes hold-to-talk release handling, reloads evicted speech models, and handles native background-time expiration.

Cleanup now rejects missing endings and excessive shortening, waits for the recovery copy to reach disk, and retains uncommitted transcripts for recovery. Rejected cleanup visibly falls back to the original transcription.

Validation: 89 focused tests passed; client-runtime typecheck passed; the native voice module and keyboard controls passed an iOS-targeted Swift typecheck against the pinned dependencies; strict Swift lint passed. Targeted JavaScript lint passed with warnings. The local mobile package typecheck reported navigation errors outside the changed voice files; CI subsequently passed the repository typecheck. Two existing Swift parameter-alignment violations in the diff viewer were corrected to unblock native CI. Physical-device long-recording verification remains outstanding; this PR does not publish an app build.

Scope: iOS/iPadOS composer microphone and hardware shortcuts, shared voice controller, local speech engines, and voice documentation. No provider adapter, server protocol, or connection-mode changes. Web, desktop, and Android do not expose this dictation implementation. UI changes are status/accessibility wording; no visual layout changes.

Implemented with GPT-6 in the Codex harness.

[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	14.2 KiB	14.0 KiB	−191 B (−1.3%)	15.1 KiB	✅
Codex	Thread snapshot wire	7.3 KiB	7.3 KiB	−4 B (−0.1%)	7.3 KiB	✅
Codex	Live turn WebSocket wire	7.0 KiB	6.8 KiB	−187 B (−2.6%)	7.8 KiB	✅
Codex	Live turn WebSocket decoded	59.8 KiB	58.4 KiB	−1.5 KiB (−2.4%)	66.4 KiB	✅
Codex	Live turn messages	10	9	−1 (−10.0%)	21	✅
Claude	Total thread wire	14.2 KiB	14.0 KiB	−163 B (−1.1%)	15.1 KiB	✅
Claude	Thread snapshot wire	7.3 KiB	7.3 KiB	+4 B (+0.1%)	7.3 KiB	✅
Claude	Live turn WebSocket wire	6.9 KiB	6.8 KiB	−167 B (−2.4%)	7.8 KiB	✅
Claude	Live turn WebSocket decoded	60.6 KiB	59.2 KiB	−1.5 KiB (−2.4%)	66.4 KiB	✅
Claude	Live turn messages	10	9	−1 (−10.0%)	21	✅

Baseline: 12b9cb0 · PR result: 36cda19 · 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: 116.2 KiB
• Claude decoded thread snapshot: 116.9 KiB

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