fix(desktop): support voice barge-in during TTS

[PrCu]

Summary

• expose first-speech detection from the microphone recorder
• keep a microphone recorder active while voice-conversation TTS is playing
• stop TTS and close the current turn when the user starts speaking

Root cause

The previous voice-conversation implementation waited for playSpeechText() to finish before starting the next recorder. As a result, users could not barge in during TTS, and the next turn depended on a later UI interaction. The recorder now observes the first speech threshold during playback and interrupts the current TTS path.

Test plan

• git diff --check passed
• source-level lifecycle review completed
• full Desktop TypeScript/lint tests are blocked in this checkout because apps/desktop/node_modules does not contain executable tsc/eslint; no dependency installation was performed
• requires Desktop manual verification for TTS interruption, consecutive turns, and microphone cleanup

[alt-glitch]

This was generated by AI during triage.

Related to #59637: both seek interruptible Desktop TTS, while this implementation specifically keeps the recorder alive during playback and routes first speech into the normal turn-close path.

[teknium1]

Thanks for targeting a real Desktop voice-conversation gap: current main waits at apps/desktop/src/app/chat/composer/hooks/use-voice-conversation.ts:228 for TTS completion before it can re-enter the recorder loop.

Problems

• apps/desktop/src/app/chat/composer/hooks/use-voice-conversation.ts:235-237 stops playback and starts handleTurn(true), but speak() then resumes and unconditionally sets pendingStartRef and status to idle at lines 245-247. The idle loop at lines 378-384 can start listening while that same recorder is still stopping/transcribing. stopVoicePlayback() resolves the active playback promise (apps/desktop/src/lib/voice-playback.ts:40-50), and playSpeechText() reports the cancellation as false at lines 134-141, but this result is currently discarded.
• The diff adds no behavior test for interruption versus normal completion.

Suggested changes

• Preserve the playSpeechText() completion result (or a barge-in ref) and bypass the normal re-arm path when playback was interrupted; let handleTurn(true) own that transition.
• Add tests for one-shot first-speech detection, no concurrent recorder/timeout during barge-in transcription, and normal TTS re-arming.

Automated hermes-sweeper review.

[teknium1]

stopVoicePlayback() resolves playSpeechText(), so this handleTurn(true) races the enclosing speak() finally block, which unconditionally marks the conversation idle and re-arms listening. Track the interruption and skip that normal-completion re-arm path so this turn exclusively owns recorder stop/transcription.

[teknium1]

Closing as superseded by the merged voice playback/barge-in rework #69511 (93e9061f15, merged 2026-07-22) plus #69602: the desktop voice stack was rewritten around speech-stream sessions with barge-in capture and native resume-listening, and the code this PR patches no longer exists. Barge-in during TTS shipped as part of #69511 — thanks for pushing this capability.

If anything in this area still misbehaves on current main, please open a fresh PR against the new stack.