Mac M1 optimizations, fix train pipeline, add Hey Cosmo wake word model

- Fix install_mac.sh: use venv + Python 3.12 (3.14 incompatible with ML libs)
- Fix run_mac.sh: activate venv, add CPU thread optimization env vars
- Fix agent.py: remove f-string from SYSTEM_PROMPT template (NameError on import)
- Add missing deps: sounddevice, pydub, imageio-ffmpeg, omegaconf
- Optimize for M1: torch.inference_mode, set_num_threads, OMP/MKL tuning
- Switch to qwen2.5:3b for faster LLM responses on Mac
- Switch Whisper to medium model with auto compute (small+int8 had poor Russian)
- Add initial_prompt for better Russian transcription
- Add open_app tool for native macOS app launching
- Fix TTS: sanitize Latin text to Cyrillic for Silero compatibility
- Fix wake word echo: add cooldown after TTS, reset model state, raise threshold
- Make "Слушаю" TTS synchronous to avoid mic interference
- Fix train Dockerfile: remove tensorflow/onnx2tf (only ONNX needed), fix deps
- Fix train.sh: use wget for dataset download, add --shm-size=2g
- Add trained hey_cosmo.onnx wake word model
- Add TODO section to CLAUDE.md (ChatterBox TTS, Ollama Modelfile ideas)

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

train_wakeword/cosmo_config.yaml

@@ -40,6 +40,11 @@ batch_n_per_class:
   "adversarial_negative": 50
   "positive": 50
 
+# Пути для аугментации (пустые — аугментация без RIR и фонового шума)
+rir_paths: []
+background_paths: []
+background_paths_duplication_rate: []
+
 # Архитектура модели
 model_type: "dnn"
 layer_size: 32