Without it faster-whisper defaults to an 80-mel feature extractor. These weights report n_mels=80, so every transcription failed with 'Invalid input features shape: expected (1,80,3000)'. feature_size is read from the CT2 model itself and verified by transcribing a real clip before upload. Weights untouched. Dr. J (Opus 5), clinic log doctor-logs/2026-08-05_stt-fleet-qa.md
add tokenizer.json -- required for offline load (faster-whisper)
