refactor: VAD upgrade, retry, dead code cleanup, AGENT removal

- audio: switch VAD to webrtcvad with RMS gate + fallback to RMS
- audio: honor FOLLOWUP_TIMEOUT — short silence wait after bot response
- llm: retry with exponential backoff on network errors and 5xx
- llm: VOICE_MAX_TOKENS env (default 300) instead of hardcoded 150
- tts: optional VAD-based barge-in (BARGE_IN_ENABLED, off by default)
- tts: remove dead start_barge_in_listener / was_barge_in helpers
- config: drop AGENT/LUSYA_AGENT — routing happens via session_key
- modes: remove unused imports, pass FOLLOWUP_TIMEOUT to follow-up record()
- docs: full rewrite of README and CLAUDE.md to match current architecture

.env.example

@@ -1,13 +1,12 @@
 # OpenClaw Gateway — Cosmo
+# Роутинг к агенту идёт через COSMO_SESSION_KEY, отдельный AGENT не нужен.
 GATEWAY_URL=http://192.168.31.103:18789
 GATEWAY_TOKEN=your_openclaw_token_here
-AGENT=openclaw/main
 VOICE_MODEL=openai/gpt-5.4-mini
 
 # OpenClaw Gateway — Люся
 LUSYA_GATEWAY_URL=http://192.168.31.103:18790
 LUSYA_GATEWAY_TOKEN=your_openclaw_token_here
-LUSYA_AGENT=openclaw/main
 LUSYA_VOICE_MODEL=openai/gpt-5.4-mini
 
 # STT (Groq)
@@ -31,6 +30,17 @@ SILENCE_THRESHOLD=500
 SILENCE_DURATION=1.5
 MAX_DURATION=15
 FOLLOWUP_TIMEOUT=8
+VAD_AGGRESSIVENESS=2       # webrtcvad 0..3, больше = строже
+
+# LLM
+VOICE_MAX_TOKENS=300
+LLM_RETRIES=3
+
+# Barge-in (прерывание TTS голосом). Работает только при разнесённых мике/колонке
+# или в наушниках — иначе собственный TTS будет триггерить прерывание.
+BARGE_IN_ENABLED=false
+BARGE_IN_THRESHOLD=1500     # RMS выше SILENCE_THRESHOLD
+BARGE_IN_WARMUP=0.8         # сек пропуска в начале TTS
 
 # Логирование
 LOG_FILE=errors.log