PR #44 merged: fix(pack): VAD-filter sysaudio transcription (silence hallucinations)

Closes #43.

Sysaudio is silent in most takes, and Whisper fed silence hallucinates loops — the last real pack's sys_transcript was "Thank you." × 7, timestamps and all.

• asr-sidecar: /transcribe gains an optional vad form flag → faster-whisper's vad_filter (Silero VAD), so silence never reaches the model. Default off — existing callers unchanged.
• AsrClient: transcribe($path, bool $vad = false), sent as a multipart field.
• ProcessPack: only the sysaudio path opts in. The mic track keeps VAD off — it's real speech end-to-end and VAD can clip quiet word edges from a verbatim take.

Test pins the wiring: mic request carries vad=false, sysaudio carries vad=true. Full suite green (147).

🤖 Generated with Claude Code