Feat: Vulkan/RADV-Engine + vocab-gepruefte Spec-Drafts + Provisioning/Sync

Engine-Cutover ROCm/HIP -> Vulkan/RADV (gfx1151): +12-22% tg auf MoE (llama-bench
verifiziert, fast 53->65 t/s). ROCm-Build bleibt als Rollback unter /opt/llamacpp.

- Backend: vocab-aware Speculative Decoding. services/gguf_meta.py liest den
  Tokenizer-Fingerprint (model/pre/n_vocab) direkt aus dem GGUF-Header (ohne Modell-Load);
  register_model + migrate_config haengen nur VOCAB-KOMPATIBLE Drafts an (inkl. --spec-type,
  das in dieser llama.cpp-Generation noetig ist). Neue Endpoints /api/models/drafts + /{id}/draft.
- Frontend: idiotensichere Spec-Draft-UI (SpecDraftModal) - nur kompatible Drafts waehlbar,
  inkompatible gesperrt mit Begruendung; SPEC/SPEC?-Badge nach echtem Aktiv-Status; Rolle in AddModel.
- maintenance.py: Engine-Update-Quelle -> ggml-org/llama.cpp (Build-Nummer-Vergleich),
  ENGINE_PATH=/opt/llamacpp-vulkan.
- Startup-Warmup der brains (deploy/warmup.sh, self-detaching ExecStartPost) + deploy/provision-engine.sh.
- Cleanup: tote LiteLLM gateway/config.yaml + alle Referenzen (config.py/backup.py/backup.sh) entfernt;
  README + docs/memory aktualisiert.

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

deploy/backup.sh

@@ -6,11 +6,10 @@ set -euo pipefail
 MODELS_DIR="${MC_MODELS_DIR:-/srv/models}"
 MEM_DB="${MC_MEMORY_DB:-$MODELS_DIR/mc2-memory.db}"
 LSWAP="${MC_CONFIG_PATH:-/etc/llama-swap/config.yaml}"
-GW="${MC_GATEWAY_CONFIG:-/opt/mission-control-2/gateway/config.yaml}"
 DEST="$MODELS_DIR/mc2-backups/$(date +%Y%m%d-%H%M%S)"
 
 mkdir -p "$DEST"
-for f in "$MEM_DB" "$MEM_DB-wal" "$MEM_DB-shm" "$LSWAP" "$GW"; do
+for f in "$MEM_DB" "$MEM_DB-wal" "$MEM_DB-shm" "$LSWAP"; do
   [ -f "$f" ] && cp -p "$f" "$DEST/" || true
 done