fix: add gcc/build-essential for webrtcvad, drop silero-vad pip install

resemblyzer depends on webrtcvad which compiles C extensions — gcc and
build-essential are required in slim containers. silero-vad pip package
has a broken native .so in python:3.12-slim; VAD model is now loaded at
runtime via torch.hub.load() which downloads on first use.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

pipeline/Dockerfile

@@ -1,20 +1,22 @@
 FROM python:3.12-slim
 
 RUN apt-get update && apt-get install -y --no-install-recommends \
-    ffmpeg libsndfile1 && rm -rf /var/lib/apt/lists/*
+    ffmpeg libsndfile1 gcc build-essential && rm -rf /var/lib/apt/lists/*
 
 WORKDIR /app
 
 COPY pyproject.toml .
+# Install torch CPU-only first (avoids pulling GPU variant), then rest of deps
+# silero-vad pip package excluded — runtime uses torch.hub.load instead
 RUN pip install --no-cache-dir torch --index-url https://download.pytorch.org/whl/cpu && \
-    pip install --no-cache-dir resemblyzer silero-vad scipy numpy && \
+    pip install --no-cache-dir resemblyzer scipy numpy && \
     pip install --no-cache-dir fastapi uvicorn[standard] pydantic-settings httpx websockets livekit-api
 
 COPY app/ app/
 
-# Pre-download VAD and resemblyzer models into the image
-RUN python -c "import torch; torch.hub.load('snakers4/silero-vad', 'silero_vad', trust_repo=True)" && \
-    python -c "from resemblyzer import VoiceEncoder; VoiceEncoder()"
+# Pre-download resemblyzer model (VoiceEncoder downloads ~25MB pretrained weights)
+RUN python -c "from resemblyzer import VoiceEncoder; VoiceEncoder()"
+# Note: silero-vad model downloads lazily at first container startup via torch.hub
 
 ENV PYTHONUNBUFFERED=1
 CMD ["python", "-m", "uvicorn", "app.pipeline.main:app", "--host", "0.0.0.0", "--port", "8300"]