refactor(llm): GPU-Architektur - Text lokal, Vision Cloud

- MODEL_VISION von qwen3-vl:32b (Ollama) auf openai/gpt-4o-mini (OpenRouter)
- Vision-Modell aus OLLAMA_MODELS entfernt (kein GPU-Swap mehr)
- keep_alive=-1 fuer Ollama: Textmodell bleibt permanent im VRAM
- RTX 3090 wird dedizierter Text+Embedding-Server
- Neue Dienste koennen Ollama mitnutzen ohne Konkurrenz

homelab-ai-bot/llm.py

@@ -19,11 +19,11 @@ OLLAMA_BASE = "http://100.84.255.83:11434"
 OPENROUTER_BASE = "https://openrouter.ai/api/v1"
 
 MODEL_LOCAL = "qwen3:30b-a3b"
-MODEL_VISION = "qwen3-vl:32b"
+MODEL_VISION = "openai/gpt-4o-mini"
 MODEL_ONLINE = "perplexity/sonar"
 FALLBACK_MODEL = "qwen2.5:14b"
 MAX_TOOL_ROUNDS = 3
-OLLAMA_MODELS = {MODEL_LOCAL, MODEL_VISION, FALLBACK_MODEL}
+OLLAMA_MODELS = {MODEL_LOCAL, FALLBACK_MODEL}
 
 PASSTHROUGH_TOOLS = {"get_temperaturen", "get_energie", "get_heizung"}
 
@@ -269,6 +269,7 @@ def _call_api(messages: list, api_key: str, use_tools: bool = True,
         headers = {"Content-Type": "application/json"}
         timeout = _ollama_timeout_for(chosen)
         _add_no_think(payload.get("messages", []))
+        payload["keep_alive"] = -1
     else:
         url = f"{OPENROUTER_BASE}/chat/completions"
         headers = {"Authorization": f"Bearer {api_key}"}