feat: VoiceCoach v2 - OpenClaw gateway + Speaches STT/TTS

- pipeline.py: completely rewritten to use Speaches STT/TTS and OpenClaw LLM
- main.py: v2 with OPENCLAW_TOKEN support
- requirements.txt: slimmed (removed anthropic, faster-whisper direct deps)
- Dockerfile: slimmer build (no model pre-download)
- whisper_stt.py, piper_tts.py: removed (replaced by Speaches API)
- docker-compose.yml: v2 with speaches + openclaw services, piper-tts removed
- openclaw-agent/system_prompt.md: new directory with agent system prompt

pipecat-agent/whisper_stt.py

@@ -1,35 +0,0 @@
-"""
-Local STT service using faster-whisper.
-Replaces Deepgram with fully self-hosted speech-to-text.
-"""
-import io
-import logging
-import wave
-import numpy as np
-from faster_whisper import WhisperModel
-
-logger = logging.getLogger(__name__)
-
-
-class LocalWhisperSTT:
-    """Wraps faster-whisper for use in the Pipecat pipeline."""
-
-    def __init__(self, model_size: str = "medium", device: str = "cpu", compute_type: str = "int8"):
-        logger.info(f"Loading faster-whisper model: {model_size} on {device}")
-        self.model = WhisperModel(model_size, device=device, compute_type=compute_type)
-        logger.info("faster-whisper model loaded.")
-
-    def transcribe(self, audio_bytes: bytes, sample_rate: int = 16000) -> str:
-        """Transcribe raw PCM audio bytes to text."""
-        audio_array = np.frombuffer(audio_bytes, dtype=np.int16).astype(np.float32) / 32768.0
-
-        segments, info = self.model.transcribe(
-            audio_array,
-            beam_size=5,
-            language="en",
-            vad_filter=True,
-            vad_parameters=dict(min_silence_duration_ms=500),
-        )
-
-        text = " ".join(seg.text.strip() for seg in segments)
-        return text