Performance
Reduced CPU usage during active transcription from ~84% to ~17% (80% reduction).
- Slow polling loop from 100ms to 250ms with dirty-checking for minibar updates
- Create minibar NSHostingView once, drive updates via @observable state
- Remove dead system audio RMS computation and per-buffer diagLog
- Replace scalar RMS loops with vDSP hardware-accelerated RMS in mic capture
- Replace O(n) VAD buffer shifts with index-based consumption
- Batch diarization audio feeding into ~1s chunks (was per-buffer Task spawn)
- Default batch refinement to off for new installs
Idle wake-ups dropped from 22,070 to 110. Thread count reduced from 32 to 15.