VadHandler is now the sole mic owner — removed separate AudioRecorder that
caused audio focus contention on Android. Audio from onSpeechEnd is encoded
as WAV for Whisper. Provider switched to autoDispose to match widget lifecycle,
preventing defunct element assertions. Recording UI deferred until mic is open.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Voice: recreate AudioRecorder each session to avoid stale native
state, improve permission handling (re-check after settings), show
feedback after 3 consecutive empty transcripts, allow STT-only mode
when TTS is unavailable.
Knowledge: hydrate IDs immediately after loading more pages so
scroll-based pagination doesn't stall at the bottom.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Three bugs causing silent STT failure on Android:
1. AudioEncoder.opus produces an OGG container on Android but the file
was named .webm — faster-whisper rejected it due to format mismatch.
Changed to AudioEncoder.aacLc + .m4a (reliable on all Android versions).
Updated voice_api.dart to send audio/mp4 MIME type accordingly.
2. onError callback was never stored, so errors in _startListening() and
_handleSilence() were silently swallowed. Now stored as _onError and
called before exitVoiceMode() on any failure.
3. Amplitude stream can emit NaN or ±Infinity on some Android devices
during recorder initialisation. NaN < -40.0 is false, so silence was
never detected. Now treated explicitly as silence.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
bvandeusen