Feat: Sprechen-Tab — mit Hermes per Sprache reden (Browser-Voice + 3D-Avatar)

Voll-Duplex Sprach-Interaktion vom lokalen PC mit dem vollen Hermes-Agenten
(api_server :8642, OpenAI-kompatibel → gleiche Tools + geteiltes Mem0 wie CLI/Telegram).

- Voice-Sidecar (voice_service/, eigenes Py3.12-venv ~/.voice, :8650): STT faster-whisper
  (medium, de) + gestuftes TTS — Piper (schnell, Default) + Chatterbox (premium, Voice-Cloning,
  lazy-load, CPU-Start). Analog mem0_service.
- Backend: routers/voice.py (Proxy /api/voice/stt|tts|voices + /chat-SSE an Hermes mit
  Bearer API_SERVER_KEY + X-Hermes-Session-Id für server-seitigen Verlauf). config.py:
  VOICE_SERVICE_URL + HERMES_API_KEY (Fallback aus ~/.hermes/.env). System-Dienstliste +
  Wartung (Restart/Logs) um voice-service ergänzt.
- Frontend: Sprechen-Tab mit 3D-Avatar (VRM via three-vrm) — Lippensync (Web-Audio-Pegel),
  Blinzeln, Sentiment-Mimik, Ruhepose. Avatar-Picker (CORS-freie Galerie + .vrm-Upload + URL
  + VRoid-Hub-Link) + Stimm-Auswahl. Push-to-talk (Knopf/Leertaste). Deps: three, r3f, drei.
- Deploy: deploy/voice-service.service + deploy.sh (idempotenter Sidecar-Install, enable, restart).

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

frontend/src/lib/voice/usePushToTalk.ts

@@ -0,0 +1,50 @@
+import { useCallback, useEffect, useRef, useState } from "react"
+
+// Push-to-talk-Aufnahme über MediaRecorder. `start` beim Drücken (Taste/Button), `stop` beim
+// Loslassen → fertiges Audio-Blob (webm/opus) geht an `onAudio`. Bewusst minimal & generisch.
+export function usePushToTalk(onAudio: (blob: Blob) => void) {
+  const [recording, setRecording] = useState(false)
+  const recRef = useRef<MediaRecorder | null>(null)
+  const chunksRef = useRef<Blob[]>([])
+  const streamRef = useRef<MediaStream | null>(null)
+
+  const start = useCallback(async () => {
+    if (recRef.current) return
+    let stream: MediaStream
+    try {
+      stream = await navigator.mediaDevices.getUserMedia({ audio: true })
+    } catch (e) {
+      console.error("Mikrofon-Zugriff verweigert:", e)
+      return
+    }
+    streamRef.current = stream
+    const mime = MediaRecorder.isTypeSupported("audio/webm;codecs=opus")
+      ? "audio/webm;codecs=opus"
+      : "audio/webm"
+    const rec = new MediaRecorder(stream, { mimeType: mime })
+    chunksRef.current = []
+    rec.ondataavailable = (e) => { if (e.data.size) chunksRef.current.push(e.data) }
+    rec.onstop = () => {
+      const blob = new Blob(chunksRef.current, { type: mime })
+      streamRef.current?.getTracks().forEach((t) => t.stop())
+      streamRef.current = null
+      recRef.current = null
+      setRecording(false)
+      if (blob.size > 1200) onAudio(blob)   // Mini-Blobs (Versehen) ignorieren
+    }
+    rec.start()
+    recRef.current = rec
+    setRecording(true)
+  }, [onAudio])
+
+  const stop = useCallback(() => {
+    recRef.current?.stop()
+  }, [])
+
+  useEffect(() => () => {
+    recRef.current?.stop()
+    streamRef.current?.getTracks().forEach((t) => t.stop())
+  }, [])
+
+  return { recording, start, stop }
+}