feat(admin): Phase 3 — Laufzeit-Konfiguration ohne Server-Neustart

Neue Tabelle config_overrides in SQLite; RuntimeSettings-Wrapper liest
überschreibbare Felder mit 30s TTL-Cache aus der DB und fällt auf
.env-Werte zurück. dependencies.py und quota.py nutzen runtime_settings
als Default statt des statischen Settings-Singletons.

16 Felder überschreibbar: STT/LLM/TTS-Provider, LLM-Modelle, Stimmen,
Systemprompt, Temperatur, Tageskontingent, Normalisierung u.a.

Backend: GET/PUT/DELETE /api/admin/config/{key}
Admin-UI: neuer Tab "⚙ Einstellungen" mit Inline-Edit und Reset.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

app/runtime_config.py

@@ -0,0 +1,97 @@
+"""Laufzeit-Konfigurationsüberschreibungen aus der Datenbank.
+
+Einzelne Settings-Felder können zur Laufzeit via Admin-UI geändert werden,
+ohne den Server neu zu starten. Die Werte liegen in der Tabelle
+`config_overrides` und werden mit 30s TTL gecacht.
+
+Nur Felder aus RUNTIME_SETTABLE sind überschreibbar — alle anderen
+kommen weiterhin aus .env / TOML / pydantic-settings.
+"""
+
+import threading
+import time
+from typing import Any
+
+from app.config import Settings, settings as _base
+
+# (label, type_str, hint)
+RUNTIME_SETTABLE: dict[str, tuple[str, str, str]] = {
+    "default_stt_provider":          ("STT-Provider (Standard)",       "str",   "openrouter | faster-whisper"),
+    "default_llm_provider":          ("LLM-Provider (Standard)",       "str",   "openrouter | local-openai-compatible"),
+    "default_tts_provider":          ("TTS-Provider (Standard)",       "str",   "openrouter | piper | chatterbox"),
+    "default_language":              ("Sprache (Standard)",            "str",   "de | en | …"),
+    "openrouter_llm_model":          ("LLM-Modell (OpenRouter)",       "str",   "z.B. google/gemini-3.1-flash-lite"),
+    "openrouter_tts_model":          ("TTS-Modell (OpenRouter)",       "str",   "z.B. google/gemini-3.1-flash-tts-preview"),
+    "openrouter_tts_voice":          ("TTS-Stimme (OpenRouter)",       "str",   "z.B. Zephyr, Puck, Kore"),
+    "piper_voice":                   ("TTS-Stimme (piper)",            "str",   "z.B. de_DE-thorsten-high"),
+    "local_llm_system_prompt":       ("Systemprompt (lokal)",          "str",   "Freier Text"),
+    "local_llm_temperature":         ("Temperatur (lokal)",            "float", "0.0–2.0"),
+    "local_llm_max_tokens":          ("Max. Tokens (lokal)",           "int",   "0 = kein Limit"),
+    "tts_normalize_level":           ("TTS-Normalisierung",            "str",   "auto | full | light | off"),
+    "audio_stream_default":          ("Audio-Streaming Standard",      "bool",  "true | false"),
+    "memory_extraction_enabled":     ("Erinnerungs-Extraktion",        "bool",  "true | false"),
+    "memory_extraction_every_n_turns": ("Extraktion alle N Turns",     "int",   "z.B. 3"),
+    "daily_request_limit":           ("Tageskontingent (global)",      "int",   "0 = unbegrenzt"),
+}
+
+_TTL = 30.0
+_cache: dict[str, str] = {}
+_cache_time: float = 0.0
+_lock = threading.Lock()
+
+
+def _coerce(key: str, raw: str) -> Any:
+    _, type_str, _ = RUNTIME_SETTABLE[key]
+    try:
+        if type_str == "bool":
+            return raw.strip().lower() in ("1", "true", "yes")
+        if type_str == "int":
+            return int(raw)
+        if type_str == "float":
+            return float(raw)
+    except (ValueError, AttributeError):
+        pass
+    return raw
+
+
+def _get_cache() -> dict[str, str]:
+    global _cache, _cache_time
+    now = time.monotonic()
+    with _lock:
+        if now - _cache_time < _TTL:
+            return _cache
+        try:
+            from app.dependencies import get_store
+            _cache = get_store().get_config_overrides()
+            _cache_time = now
+        except Exception:
+            pass
+        return _cache
+
+
+def invalidate_cache() -> None:
+    global _cache_time
+    with _lock:
+        _cache_time = 0.0
+
+
+class RuntimeSettings:
+    """Wraps Settings; liest überschreibbare Felder aus der DB (30s TTL)."""
+
+    def __init__(self, base: Settings):
+        object.__setattr__(self, "_base", base)
+
+    def __getattr__(self, name: str) -> Any:
+        if name in RUNTIME_SETTABLE:
+            overrides = _get_cache()
+            if name in overrides:
+                return _coerce(name, overrides[name])
+        return getattr(object.__getattribute__(self, "_base"), name)
+
+    # Delegiere Pydantic-Metadaten ans Basis-Objekt.
+    @property
+    def model_fields(self):
+        return self._base.model_fields
+
+
+runtime_settings = RuntimeSettings(_base)