feat(dross): voice Phase 2a — local whisper transcribe + mic (2.12.0)

faster-whisper (small.en, GPU+CPU fallback) on CT 102 → POST
/api/voice/transcribe (multer→whisper client) → mic in the bubble
records (MediaRecorder), uploads, drops the transcript into the input
to review-and-send. Infra scripts in deploy/whisper/. Retention (P2b)
next. NOTE: mic needs a secure context (the https domain), not the LAN IP.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

deploy/whisper/README.md

@@ -0,0 +1,16 @@
+# faster-whisper service (Dross voice STT)
+
+Runs on **CT 102** (the Ollama box, `192.168.1.185`), bare-metal (no Docker), on the
+RTX A2000 with CPU fallback. OpenAI-style `/transcribe` consumed by void-app
+`lib/voice/whisper.js` (`WHISPER_URL=http://192.168.1.185:8001`).
+
+## Install (on CT 102)
+```
+scp deploy/whisper/{server.py,setup.sh} root@192.168.1.185:/opt/whisper_server.py /root/setup.sh
+ssh root@192.168.1.185 'bash /root/setup.sh && install -m644 /opt/whisper_server.py /opt/whisper/server.py && systemctl enable --now whisper'
+curl http://192.168.1.185:8001/health   # {"ok":true,"model":"small.en","device":"cuda"}
+```
+- venv at `/opt/whisper/venv`; model `small.en` (env `WHISPER_MODEL`); CUDA libs via
+  `nvidia-cublas-cu12`/`nvidia-cudnn-cu12` pip wheels (LD_LIBRARY_PATH in the unit).
+- GPU → CPU fallback is in `server.py` `load()`.
+- **CT 102 disk was expanded +20G** (was 89% full) before install.

deploy/whisper/server.py

@@ -0,0 +1,35 @@
+import os, tempfile
+from fastapi import FastAPI, UploadFile, File, HTTPException
+from faster_whisper import WhisperModel
+
+MODEL = os.environ.get("WHISPER_MODEL", "small.en")
+app = FastAPI()
+model = None
+device_used = None
+
+def load():
+    global model, device_used
+    try:
+        model = WhisperModel(MODEL, device="cuda", compute_type="int8_float16")
+        device_used = "cuda"
+    except Exception:
+        model = WhisperModel(MODEL, device="cpu", compute_type="int8")
+        device_used = "cpu"
+
+load()
+
+@app.get("/health")
+def health():
+    return {"ok": True, "model": MODEL, "device": device_used}
+
+@app.post("/transcribe")
+async def transcribe(file: UploadFile = File(...)):
+    data = await file.read()
+    if not data:
+        raise HTTPException(400, "empty audio")
+    with tempfile.NamedTemporaryFile(suffix=".bin") as f:
+        f.write(data); f.flush()
+        segments, info = model.transcribe(f.name, beam_size=1, vad_filter=True)
+        text = "".join(s.text for s in segments).strip()
+    return {"text": text, "language": info.language,
+            "duration": round(info.duration, 2), "device": device_used}

deploy/whisper/setup.sh

@@ -0,0 +1,26 @@
+#!/usr/bin/env bash
+set -e
+export DEBIAN_FRONTEND=noninteractive
+apt-get update -qq
+apt-get install -y -qq python3-pip python3-venv ffmpeg >/dev/null
+mkdir -p /opt/whisper
+python3 -m venv /opt/whisper/venv
+/opt/whisper/venv/bin/pip install -q --upgrade pip
+/opt/whisper/venv/bin/pip install -q faster-whisper fastapi "uvicorn[standard]" python-multipart nvidia-cublas-cu12 nvidia-cudnn-cu12
+SITE=/opt/whisper/venv/lib/python3.12/site-packages
+cat > /etc/systemd/system/whisper.service <<UNIT
+[Unit]
+Description=faster-whisper transcription server (Dross voice)
+After=network.target
+[Service]
+Type=simple
+WorkingDirectory=/opt/whisper
+Environment=WHISPER_MODEL=small.en
+Environment=LD_LIBRARY_PATH=${SITE}/nvidia/cublas/lib:${SITE}/nvidia/cudnn/lib
+ExecStart=/opt/whisper/venv/bin/uvicorn server:app --host 0.0.0.0 --port 8001
+Restart=on-failure
+[Install]
+WantedBy=multi-user.target
+UNIT
+systemctl daemon-reload
+echo "deps+unit installed"