feat(admin): LLM-/GPU-Status-Karte (read-only) — Plan-Schritt 2

- app/admin_llm.py: read-only Statusabfragen ohne sudo (docker ps, ollama ps,
  nvidia-smi, systemctl --user) mit sicheren Defaults bei fehlenden Tools.
- GET /api/admin/llm/status (require_admin): Backend, Modell, Backend-Status,
  geladene Ollama-Modelle, GPU-Auslastung, Gateway-Dienst-Status.
- Admin Status-Tab: LLM-Backend-Karte mit GPU-Balken.
- Tests: Auth-Gate + Antwortschema (160 grün).
- Doku §7.5: Status-Tab um LLM/GPU-Karte ergänzt.

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

app/admin_llm.py

@@ -0,0 +1,115 @@
+"""Read-only Statusabfragen rund um das lokale LLM-Backend (fuer das Admin-Panel).
+
+Alles nur lesend, ohne sudo: `docker ps`, `ollama ps`, `nvidia-smi`, `systemctl --user`.
+Fehlende Tools oder Fehler fuehren zu sicheren Defaults (None/[]), nie zu Exceptions.
+"""
+
+from __future__ import annotations
+
+import asyncio
+import os
+import shutil
+
+from app.config import settings
+
+
+async def _run(cmd: list[str], timeout: float = 6.0) -> str | None:
+    """Fuehrt ein Kommando aus (shell=False) und liefert stdout, oder None bei Fehler."""
+    if not shutil.which(cmd[0]):
+        return None
+    try:
+        proc = await asyncio.create_subprocess_exec(
+            *cmd,
+            stdout=asyncio.subprocess.PIPE,
+            stderr=asyncio.subprocess.DEVNULL,
+            env={**os.environ, "XDG_RUNTIME_DIR": os.environ.get(
+                "XDG_RUNTIME_DIR", f"/run/user/{os.getuid()}")},
+        )
+        out, _ = await asyncio.wait_for(proc.communicate(), timeout=timeout)
+        if proc.returncode != 0:
+            return None
+        return out.decode("utf-8", "replace")
+    except (asyncio.TimeoutError, OSError):
+        return None
+
+
+def _detect_backend(base_url: str) -> str:
+    if ":11434" in base_url:
+        return "ollama"
+    if ":8001" in base_url:
+        return "llamacpp"
+    return "unknown"
+
+
+async def _llamacpp_running(container: str = "va_llm") -> bool:
+    out = await _run(["docker", "ps", "--filter", f"name=^{container}$", "--format", "{{.Names}}"])
+    return bool(out and container in out)
+
+
+async def _ollama_loaded() -> tuple[bool, list[dict]]:
+    """(dienst_erreichbar, [geladene Modelle])."""
+    out = await _run(["ollama", "ps"])
+    if out is None:
+        return False, []
+    models: list[dict] = []
+    lines = [ln for ln in out.splitlines() if ln.strip()]
+    for ln in lines[1:]:  # Kopfzeile ueberspringen
+        # Spalten sind durch 2+ Leerzeichen getrennt: NAME ID SIZE PROCESSOR CONTEXT UNTIL
+        import re
+        cols = re.split(r"\s{2,}", ln.strip())
+        if cols:
+            models.append({
+                "name": cols[0],
+                "size": cols[2] if len(cols) > 2 else "",
+                "processor": cols[3] if len(cols) > 3 else "",
+            })
+    return True, models
+
+
+async def _gpus() -> list[dict]:
+    out = await _run([
+        "nvidia-smi",
+        "--query-gpu=index,memory.used,memory.total",
+        "--format=csv,noheader,nounits",
+    ])
+    if out is None:
+        return []
+    gpus: list[dict] = []
+    for ln in out.splitlines():
+        parts = [p.strip() for p in ln.split(",")]
+        if len(parts) == 3 and parts[0].isdigit():
+            used, total = int(parts[1]), int(parts[2])
+            gpus.append({
+                "index": int(parts[0]),
+                "used_mib": used,
+                "total_mib": total,
+                "percent": round(used / total * 100) if total else 0,
+            })
+    return gpus
+
+
+async def _gateway_service_active() -> bool:
+    out = await _run(["systemctl", "--user", "is-active", "voice-assistant.service"], timeout=4.0)
+    return bool(out and out.strip() == "active")
+
+
+async def llm_status() -> dict:
+    """Aggregierter, read-only LLM-/System-Status fuer das Admin-Panel."""
+    base_url = settings.local_llm_base_url
+    backend = _detect_backend(base_url)
+    llamacpp, (ollama_reachable, ollama_models), gpus, gw = await asyncio.gather(
+        _llamacpp_running(),
+        _ollama_loaded(),
+        _gpus(),
+        _gateway_service_active(),
+    )
+    return {
+        "backend": backend,
+        "model": settings.local_llm_model,
+        "base_url": base_url,
+        "llamacpp_running": llamacpp,
+        "ollama_reachable": ollama_reachable,
+        "ollama_loaded": ollama_models,
+        "gpus": gpus,
+        "gateway_service_active": gw,
+    }