[AGENT] Improve live voice responsiveness and TTS preemption

[BASIC-BIT]

[AGENT]

Summary

Improve live voice command responsiveness and TTS playback behavior during Discord meetings.

Current behavior and likely causes

• Live voice commands depend on transcribed snippets, so transcription backlog can delay command detection.
• Slow command handling can be skipped when a fast transcript already ran, even if the slow transcript would be more accurate.
• Voice-stop commands can therefore arrive minutes late in long or busy meetings.
• TTS has queue clearing support but no true user-speech barge-in/preemption path.
• TTS can feel stale or interrupted awkwardly when transcript backlog and playback queue state drift apart.

Proposed scope

• Add observability for live voice command latency from speech end to action.
• Consider a priority command-detection path separate from normal transcription backlog.
• Revisit whether slow transcripts should get a second command-detection chance for likely command phrases.
• Add a clear TTS interruption policy for user speech, high-priority commands, and confirmation prompts.
• Ensure generated-but-not-yet-playing TTS can be canceled when the queue is stopped.
• Add tests around stop command latency logic and TTS queue cancellation/preemption.

Acceptance criteria

1. Live voice command latency is measurable.
2. Stop/end meeting commands are not blocked behind normal transcript backlog when a priority path is available.
3. Slow transcription can recover command detection when fast transcription missed or misheard the command.
4. TTS playback has an explicit barge-in/preemption policy.
5. Tests cover delayed transcription, fast/slow transcript command handling, and TTS cancellation.