fix: increase LLM summarize timeout to 120s for remote cf-orch routing

20s was too tight for first-request model swaps in Ollama (model cold load
can take 30-60s). 120s matches coordinator inference timeout.

app/services/llm.py

@@ -38,7 +38,7 @@ def summarize(
     llm_url: str,
     llm_model: str,
     api_key: str | None = None,
-    timeout: float = 20.0,
+    timeout: float = 120.0,
 ) -> str | None:
     if not entries:
         return None