feat(speaker-id): Phase 1 — SpeechBrain ECAPA-TDNN Backend in whisper-bridge

Speaker-ID-Modul (Hermes-Style „echtes Gespraech ohne Wake-Word"-Vision,
Phase 1 von 5). Erkennt Stefans Stimme via 192-dim Embedding + Cosine-
Match gegen einen persistierten Fingerprint.

Module:
- speaker_id.py: lazy-loaded ECAPA-TDNN (HuggingFace), enroll/verify/
  status/delete. Fingerprint = L2-normalisierter Mittelwert aus N
  Enrollment-Samples in /voice-id/fingerprint.json.
  Fail-open: kein Fingerprint → verify() returnt (True, 0.0).
- bridge.py: 3 Message-Handler — voice_id_status_request,
  voice_id_enroll_request (samples[]: base64 16kHz int16 PCM),
  voice_id_delete_request. Enrollment laeuft im Executor (Torch
  blockt sonst die Event-Loop).
- Dockerfile: torch 2.3.1 + torchaudio mit CUDA-12.1-Wheels (sonst
  zieht speechbrain CPU-only Torch rein). Container ~1 GB groesser.
- docker-compose.yml: ./voice-id:/voice-id Bind-Mount fuer Fingerprint-
  Persistenz (ueberlebt Container-Restart).
- rvs/server.js: 6 neue Message-Types in ALLOWED_TYPES.

Phase 2 (next): App-Enrollment-Flow + Diagnostic-Voice-ID-Section.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

rvs/server.js

@@ -42,6 +42,12 @@ const ALLOWED_TYPES = new Set([
   // die feuert stt_endpoint mit dem finalen Text — kein Audio-Roundtrip.
   "stt_stream_start", "stt_audio_chunk", "stt_stream_end",
   "stt_partial", "stt_endpoint", "stt_stream_done",
+  // Speaker-ID / Voice-Enrollment (Phase 1+2): App schickt 5-10 Samples zur
+  // whisper-bridge, die berechnet einen Voice-Fingerprint (Embedding-Vektor)
+  // und nutzt ihn um nur Stefans Stimme an Whisper STT durchzulassen.
+  "voice_id_status_request", "voice_id_status_response",
+  "voice_id_enroll_request", "voice_id_enroll_response",
+  "voice_id_delete_request", "voice_id_delete_response",
   // File-Versioning (Datei-Manager in App): Versionen pro Datei listen,
   // alte Versionen herunterladen, Restore = non-destructive neuer Commit.
   "file_version_list_request", "file_version_list_response",