v0.325.220

Support multiple concurrent TCP transcriptions sharing a single loaded model

TCP server now spawns a thread per connection, each getting an independent
pipeline (~7.3MB) while sharing the model weights in VRAM/RAM. Local audio
capture (--trigger) runs alongside TCP in a background thread. Output
simplified to raw text stream (no timestamp framing). CoreML encoder caches
moved from shared model struct to per-pipeline for thread safety. --input
flag deprecated in favour of --trigger for enabling local mode.

Co-Authored-By: Claude Opus 4.6 (1M context) noreply@anthropic.com