fix(voice): keep worker protocol on a clean fd so Kokoro load noise can't corrupt READY

fix(voice): run Kokoro in a CUDA-free subprocess so it can't poison ZeroGPU

fix(voice): hide CUDA while Kokoro runs so the CPU voice doesn't poison ZeroGPU

diag(voice): HOLLOW_VOICE_OFF kill-switch to test if voice poisons ZeroGPU CUDA

fix(voice): force Kokoro onto CPU so TTS works on ZeroGPU (cuda/cpu device mismatch)

tune(finale): snappier recital — shorter inter-line breath + slightly faster whisper

fix: strip dashes from TTS input (Kokoro was saying 'dash')

fix: cap CPU math threads so Kokoro loads under memory pressure

feat: presence polish (voice every reply, idle anger, mute, fog-veil, faster finale) + race fixes

feat: voice.py — guarded Kokoro whisper with locked DSP