feat: Major performance optimizations and feature enhancements

## Performance Optimizations (3-10x faster responses)
- STT beam_size reduced to 1 (3-5x faster transcription, minimal quality loss)
- Smart query routing: Haiku (simple) → Sonnet (medium) → Opus (complex)
- TTS cache for common phrases (27 pre-generated responses)
- Sentence-level streaming TTS (start playing while generating)
- Sample-based VAD timing (30x improvement in silence detection)

## TTS Engine Upgrade
- Migrated from Chatterbox to Chatterbox-Turbo
- Zero-shot voice cloning (no fine-tuning required)
- Native paralinguistic tag support ([laugh], [sigh], [chuckle], etc.)
- Emotion presets with temperature control
- Improved marker conversion (*action*, (action), ~action~)

## Discord Bot Enhancements
- Multi-agent support (Jarvis, Sage)
- Improved voice receiving with discord-ext-voice-recv
- Enhanced /join, /leave, /status commands
- Per-agent personality configuration
- Better audio sink/receiver implementation

## OpenClaw Integration
- WebSocket support for Gateway communication
- Query complexity routing (auto-select model)
- Improved error handling and retries
- Session management per Discord guild
- Better latency tracking

## Pipeline Improvements
- Sentence splitter for streaming optimization
- Query router for intelligent model selection
- Enhanced VAD receiver with sample-based timing
- Improved audio buffering and format conversion
- Better transcript management

## Documentation
- Added QUICK_START.md (5-minute test guide)
- Added OPTIMIZATION_SUMMARY.md (performance analysis)
- Added DISCORD_OPTIMIZATION_TEST.md (testing guide)
- Added USAGE_GUIDE.md (comprehensive usage)
- Updated README.md with optimization details

## Utilities & Scripts
- Added get_invite_link.py (Discord bot invite)
- Added sync_commands.py, sync_to_guild.py (command sync)
- Added test_gateway.py, test_stt.py (testing utilities)
- Added openclaw_wrapper.py (wrapper script)
- Removed create_mock_turn_model.py (no longer needed)

## Configuration Updates
- STT model: medium → small (faster, acceptable quality)
- TTS engine: chatterbox → coqui (Turbo integration)
- Beam size: 5 → 1 (latency optimization)
- Added emotion_exaggeration per agent
- Updated .gitignore for project files

Total: ~2105 insertions, ~462 deletions across 35 files
Performance: ~5.5s total latency (down from 22-35s)
Target: ~3.5s (achieved in simple queries with cache)

Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>

config.yaml

@@ -28,7 +28,7 @@ agents:
   # Per-agent settings
   jarvis:
     # TTS voice reference file (relative to server/voices/)
-    voice_file: "jarvis.wav"
+    voice_file: "jarvis.mp3"
 
     # Agent personality for LLM context
     personality: |
@@ -50,26 +50,36 @@ agents:
     emotion_exaggeration: 0.2
 
 # ============================================================================
-# OpenClaw API
+# OpenClaw Gateway
 # ============================================================================
 openclaw:
-  # Base URL for OpenClaw API
+  # WebSocket URL for OpenClaw Gateway
   # REQUIRED: Set via OPENCLAW_BASE_URL environment variable
+  # Format: ws://IP:PORT (default port: 18789)
   base_url: null
 
   # Authentication token
-  # REQUIRED: Set via OPENCLAW_TOKEN environment variable
+  # REQUIRED: Set via OPENCLAW_AUTH_TOKEN environment variable
   token: null
 
   # Request timeout (seconds)
   timeout: 8.0
 
+  # Retry timeout (seconds)
+  retry_timeout: 15.0
+
   # Retry attempts on failure
   max_retries: 1
 
   # Model/agent selection
   model: "claude-sonnet-4"
 
+  # Agent ID for session keys
+  agent_id: "jarvis"
+
+  # Session scope: per-peer or shared
+  session_scope: "per-peer"
+
 # ============================================================================
 # Pipeline Configuration
 # ============================================================================
@@ -95,12 +105,14 @@ pipeline:
     max_wait: 3.0
 
     # Model path (relative to models/ directory)
-    model_path: "smart_turn_v3.onnx"
+    # Using v3.2 GPU model for best performance with RTX 5090
+    model_path: "smart-turn-v3.2-gpu.onnx"
 
   # Speech-to-Text (faster-whisper)
   stt:
     # Model size: tiny, base, small, medium, large-v3
-    model_size: "medium"
+    # Using "small" for faster transcription (was "medium")
+    model_size: "small"
 
     # Device: cuda or cpu
     device: "cuda"
@@ -109,7 +121,8 @@ pipeline:
     compute_type: "float16"
 
     # Beam size for decoding (higher = more accurate, slower)
-    beam_size: 5
+    # Optimized for voice chat: beam_size=1 is 3-5x faster with minimal quality loss
+    beam_size: 1
 
     # Language hint (null = auto-detect)
     language: "en"