feat(menu): nicht nutzbare TTS-Provider ausblenden + GPU-Erkennung

Menue zeigt nur tatsaechlich nutzbare Vorlese-Optionen, um Verwirrung durch
folgenlose Eintraege zu vermeiden:
- Backend ist alleinige Quelle: /api/config liefert fertiges tts_menu
  (Chatterbox nur wenn erreichbar, Cloud nur mit OpenRouter-TTS, Cartesia nur
  mit Key). piper bleibt garantierter, immer sichtbarer Default.
- Frontend blendet aus statt auszugrauen; 'Im Geraet' nur auf Mobilgeraeten
  (isMobile). Aktive/gespeicherte Auswahl auf einen sichtbaren Provider
  geklemmt -> kein Server-422.
- 4x refresh*Preset + 3x /api/config-Fetch -> 1x refreshTtsMenu; /api/config
  serverseitig 30s gecacht (spart die bisher kostenpflichtige OpenRouter-Probe
  pro Seitenladen).

GPU-Erkennung zentralisiert (keine Dopplung):
- app/capabilities.py: detect_physical_gpus/resolve_gpus (nvidia-smi).
- Neues Setting GPUS (auto|none|0,1,2), beim Start an CUDA_VISIBLE_DEVICES
  gekoppelt -> angezeigt = real nutzbar. admin_llm nutzt dieselbe Erkennung.

Tests: tests/test_capabilities.py (Maske, Kopplung, Menue-Sichtbarkeit, GPUs).

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

tests/test_capabilities.py

@@ -0,0 +1,117 @@
+"""GPU-Erkennung/-Maske und das daraus abgeleitete TTS-Menü in /api/config (offline)."""
+
+import asyncio
+import os
+
+from fastapi.testclient import TestClient
+
+import app.capabilities as cap
+import app.api.config as cfgmod
+from app.config import settings, _parse_gpu_indices, _apply_gpu_visibility
+from app.main import app
+
+client = TestClient(app)
+
+
+def _gpus(*indices):
+    async def fake():
+        return [
+            {"index": i, "name": f"GPU{i}", "used_mib": 0, "total_mib": 1, "percent": 0}
+            for i in indices
+        ]
+    return fake
+
+
+# ── Maske/Parsing ────────────────────────────────────────────────────────────
+
+def test_parse_gpu_indices():
+    assert _parse_gpu_indices("0, 2 ,1") == [0, 1, 2]
+    assert _parse_gpu_indices("") == []
+    assert _parse_gpu_indices("auto") == []
+
+
+def test_apply_gpu_visibility_none(monkeypatch):
+    monkeypatch.setattr(settings, "gpus", "none")
+    monkeypatch.delenv("CUDA_VISIBLE_DEVICES", raising=False)
+    _apply_gpu_visibility()
+    assert os.environ["CUDA_VISIBLE_DEVICES"] == ""
+
+
+def test_apply_gpu_visibility_explicit_sets_cuda(monkeypatch):
+    monkeypatch.setattr(settings, "gpus", "2,0")
+    monkeypatch.delenv("CUDA_VISIBLE_DEVICES", raising=False)
+    _apply_gpu_visibility()
+    assert os.environ["CUDA_VISIBLE_DEVICES"] == "0,2"
+
+
+def test_apply_gpu_visibility_auto_keeps_existing(monkeypatch):
+    monkeypatch.setattr(settings, "gpus", "auto")
+    monkeypatch.setenv("CUDA_VISIBLE_DEVICES", "1")
+    _apply_gpu_visibility()
+    assert os.environ["CUDA_VISIBLE_DEVICES"] == "1"  # unveraendert
+
+
+# ── resolve_gpus ─────────────────────────────────────────────────────────────
+
+def test_resolve_gpus_none(monkeypatch):
+    monkeypatch.setattr(settings, "gpus", "none")
+    assert asyncio.run(cap.resolve_gpus()) == []
+
+
+def test_resolve_gpus_auto_all(monkeypatch):
+    monkeypatch.setattr(settings, "gpus", "auto")
+    monkeypatch.delenv("CUDA_VISIBLE_DEVICES", raising=False)
+    monkeypatch.setattr(cap, "detect_physical_gpus", _gpus(0, 1))
+    assert [g["index"] for g in asyncio.run(cap.resolve_gpus())] == [0, 1]
+
+
+def test_resolve_gpus_explicit_subset(monkeypatch):
+    monkeypatch.setattr(settings, "gpus", "0,2")
+    monkeypatch.delenv("CUDA_VISIBLE_DEVICES", raising=False)
+    monkeypatch.setattr(cap, "detect_physical_gpus", _gpus(0, 1, 2))
+    assert [g["index"] for g in asyncio.run(cap.resolve_gpus())] == [0, 2]
+
+
+def test_resolve_gpus_intersects_cuda_mask(monkeypatch):
+    monkeypatch.setattr(settings, "gpus", "auto")
+    monkeypatch.setenv("CUDA_VISIBLE_DEVICES", "1")
+    monkeypatch.setattr(cap, "detect_physical_gpus", _gpus(0, 1))
+    assert [g["index"] for g in asyncio.run(cap.resolve_gpus())] == [1]
+
+
+# ── /api/config: Menü-Sichtbarkeit + gpus ────────────────────────────────────
+
+def _patch_probes(monkeypatch, *, chatterbox, openrouter, cartesia, gpus):
+    async def _t():
+        return True
+    async def _f():
+        return False
+    async def _gpu():
+        return [{"index": i, "name": f"GPU{i}"} for i in gpus]
+    monkeypatch.setattr(cfgmod, "_chatterbox_reachable", _t if chatterbox else _f)
+    monkeypatch.setattr(cfgmod, "_openrouter_tts_available", _t if openrouter else _f)
+    monkeypatch.setattr(cfgmod, "_cartesia_available", _t if cartesia else _f)
+    monkeypatch.setattr(cfgmod, "resolve_gpus", _gpu)
+    monkeypatch.setattr(cfgmod, "_CACHE", None)  # Cache je Test zuruecksetzen
+    monkeypatch.setattr(cfgmod, "_CACHE_TS", 0.0)
+
+
+def test_config_tts_menu_hides_unavailable(monkeypatch):
+    _patch_probes(monkeypatch, chatterbox=False, openrouter=False, cartesia=True, gpus=[])
+    data = client.get("/api/config").json()
+    menu = {m["provider"]: m for m in data["tts_menu"]}
+    assert menu["piper"]["visible"] is True          # garantierter Default
+    assert menu["chatterbox"]["visible"] is False    # nicht erreichbar -> aus
+    assert menu["openrouter"]["visible"] is False    # kein TTS-Modell -> aus
+    assert menu["cartesia"]["visible"] is True
+    assert menu["device"]["client_only"] is True     # Mobil-Regel macht das Frontend
+    assert data["available"]["gpus"] == []
+
+
+def test_config_reports_gpus(monkeypatch):
+    _patch_probes(monkeypatch, chatterbox=True, openrouter=True, cartesia=True, gpus=[0, 1])
+    data = client.get("/api/config").json()
+    assert [g["index"] for g in data["available"]["gpus"]] == [0, 1]
+    assert {m["provider"] for m in data["tts_menu"]} == {
+        "piper", "chatterbox", "cartesia", "openrouter", "device"
+    }