feat(voice): F010 push-to-talk on-device STT to chat

[nimat-dev]

Summary

This PR implements F010: Push-to-talk, on-device STT → chat, the first feature in Phase 04 (Voice — thin).

Changes Included:

1. On-Device STT Provider Architecture (@shellmind/mobile/src/voice):
   • types.ts: Defines ISpeechToTextProvider interface with contract:
     • isAvailable(): Promise<boolean>
     • requestPermission(): Promise<"granted" | "denied" | "undetermined">
     • startRecording(onInterimResult?: (text: string) => void): Promise<void>
     • stopRecording(): Promise<string>
     • cancelRecording(): Promise<void>
     • isRecording(): boolean
   • mock.ts: MockSpeechToTextProvider providing deterministic audio simulation, configurable fixture text, interim transcript streaming (word-by-word with delay), cancellation, and error/permission testing support.
   • native.ts: NativeSpeechToTextProvider integrating with native speech recognition (@react-native-voice/voice / Web Speech API) and falling back gracefully if native recognition is unavailable.
   • registry.ts: Provider registration with getSpeechToTextProvider(), setSpeechToTextProvider(), and resetSpeechToTextProvider().
   • Exported through packages/mobile/src/voice/index.ts and packages/mobile/src/index.ts.
2. Push-to-Talk Chat UI Integration (@shellmind/mobile/src/components/ChatScreen.tsx):
   • Push-to-talk microphone button (testID="mic-button").
   • Active recording state banner & indicator (testID="recording-indicator") with pulse label and interim transcript preview.
   • Cancel recording button (testID="voice-cancel-button").
   • Injects recognized transcript directly into testID="chat-input-field" so text is clearly visible and user-editable before sending.
   • Permission denial handling with informative banner (testID="voice-error-banner") and dismissal button (testID="voice-error-dismiss"), falling back safely to keyboard typing.
3. Tests & Evidence:
   • 7 new comprehensive mobile tests in packages/mobile/src/mobile.test.ts covering STT lifecycle, interim streaming, recording cancellation, permission denial, native fallback, registry overrides, and ChatScreen voice props.
   • 120/120 tests passing monorepo-wide (29 protocol, 52 agent, 39 mobile).
   • Clean architecture verified with dependency-cruiser (75 modules, 220 dependencies cruised, 0 violations).
   • Maestro E2E flow in .maestro/voice_stt_flow.yaml.
   • Architecture summary, test summary, and E2E trace stored in .harness/evidence/F010/.

[nimat-dev]

Maker-Checker Review: F010 (Push-to-talk, on-device STT → chat)

1. Acceptance Criteria Verification

• On-device STT provider abstraction ISpeechToTextProvider defined in packages/mobile/src/voice/types.ts.
• Mock provider MockSpeechToTextProvider supports deterministic test fixtures, interim streaming, permission overrides, and cancellation.
• Native provider NativeSpeechToTextProvider bridges native speech recognizers with graceful fallback.
• Provider registry in packages/mobile/src/voice/registry.ts supports runtime swapping.
• Push-to-talk mic button (testID="mic-button"), recording indicator (testID="recording-indicator"), cancel button (testID="voice-cancel-button"), and error banner (testID="voice-error-banner") implemented in ChatScreen.tsx.
• Speech output populates testID="chat-input-field" allowing review and editing before dispatch.
• Permission denial falls back cleanly to typing without application crashes.
• 120/120 tests pass across all packages (39 mobile tests).
• Dependency cruiser reports 0 violations across 75 modules.
• Pure core invariant preserved: voice STT is mobile-only; protocol and agent remain audio-agnostic.
• Maestro E2E specification in .maestro/voice_stt_flow.yaml.
• Harness docs and evidence logged in .harness/evidence/F010/.

2. Evaluation Scores

• Acceptance: 5/5
• Correctness: 5/5
• Boundaries: 5/5
• Modularity: 5/5
• Evidence: 5/5
• Average: 5.0 (PASS)

3. Decision

APPROVE. Ready for squash merge to main.