fix: latency-based routing to prefer Haiku over unreliable Gemini fallback

simple-shuffle randomly picked Gemini (503 overloaded) before trying Haiku.
latency-based-routing prefers the fastest responding model, with proper fallback.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: CreatmanCEO

backend/services/llm_router.py

@@ -61,9 +61,10 @@ def create_router() -> Router:
 
     return Router(
         model_list=model_list,
-        routing_strategy="simple-shuffle",
-        num_retries=2,
-        timeout=30,
+        routing_strategy="latency-based-routing",
+        num_retries=3,
+        timeout=45,
+        allowed_fails=1,
     )