fix: decouple microphone recording rate from TTS sample rate

[GrowDev1]

Fixes #491.

Summary

SAMPLE_RATE (voice_mode/config.py, 24000) was a single hardcoded constant used for both TTS playback and microphone recording. 24kHz is correct for TTS (it's what the providers emit), but wrong as a hardcoded mic capture rate — a microphone's native rate is whatever the hardware is, commonly 44.1kHz or 48kHz for USB mics, never 24kHz. Requesting a rate the device doesn't natively support isn't universally safe: some audio backends resample transparently, some don't, producing corrupted/aliased recordings on affected hardware/OS/driver combinations — exactly what #491 reports with a 48kHz-native mic.

The fix

Added RECORDING_SAMPLE_RATE = int(os.getenv("VOICEMODE_RECORDING_SAMPLE_RATE", str(SAMPLE_RATE))) — a new, independent config constant. Defaults to SAMPLE_RATE, so behavior is unchanged unless a user explicitly opts in.

Switched every recording-side usage in voice_mode/tools/converse.py to the new constant:

• prepare_audio_for_stt()'s AudioSegment construction + the Whisper-downsample comparison
• The full-quality WAV archival write, and the no-speech WAV dump
• record_audio()'s sd.rec() call, sample-count math, and debug/error logging
• record_audio_with_silence_detection()'s VAD chunk-size math, sd.InputStream() call, and the VAD-downsampling resample-length math
• The STT stats/telemetry dict (sample_rate_hz, bitrate_kbps)

All TTS/playback-side usage (all of streaming.py, core.py's chime-playback functions) is untouched — verified by checking every remaining SAMPLE_RATE reference in the repo is genuinely playback-side, not a missed recording site.

Docs fix, found while verifying this: docs/reference/environment.md and docs/guides/configuration.md already documented a VOICEMODE_SAMPLE_RATE variable — including recommending it as a fix for exactly this kind of mic-rate mismatch — but the code never actually reads that variable anywhere (confirmed via a repo-wide grep). It was a documented no-op that would have silently done nothing for anyone following the docs. Updated those entries (and one in docs/.archive/troubleshooting/wsl2-microphone-access.md) to point at the new, working VOICEMODE_RECORDING_SAMPLE_RATE instead.

Test plan

• tests/test_recording_sample_rate_config.py — new, 3 tests: default equals SAMPLE_RATE (24000, unchanged behavior), override is decoupled from TTS rate, no cross-test pollution
• tests/test_recording_sample_rate_prepare_audio.py — new, 5 tests:
  • prepare_audio_for_stt() produces WAV bytes whose actual duration is correct at an overridden rate (a real check via wave module on real exported bytes — this would fail if frame_rate weren't threaded correctly, unlike a mocked-call-args check)
  • Same at the default rate (regression guard)
  • record_audio() requests the overridden rate from sounddevice
  • VAD resample math lands on an exact 480-sample (30ms/16kHz) frame across every standard rate (8k/16k/24k/32k/44.1k/48kHz) — WebRTC VAD requires an exact frame size or raises, so this guards the most delicate part of the change
• Full existing suite (tests/test_stt_audio_saving.py, test_stt_local_compression_skip.py, test_silence_detection.py, test_vad_aggressiveness.py, test_history_buffer.py, test_audio_format_config.py): 61 passed, 2 failed (confirmed pre-existing on unmodified master too — missing ffmpeg in my environment, unrelated to this change), 5 skipped (pre-existing, unrelated)