chore(llm): swap model_candidates order — Qwen2.5-3B first, Phi-4-mini fallback

Phi-4-mini's cached modeling_phi3.py imports SlidingWindowCache which
was removed in transformers 5.x. Qwen2.5-3B uses built-in qwen2 arch
and works cleanly. Reorder so Qwen is tried first.

config/llm.yaml

@@ -28,9 +28,9 @@ backends:
     type: openai_compat
   ollama_research:
     api_key: ollama
-    base_url: http://host.docker.internal:11434/v1
+    base_url: http://ollama_research:11434/v1
     enabled: true
-    model: llama3.2:3b
+    model: llama3.1:8b
     supports_images: false
     type: openai_compat
   vision_service:
@@ -40,14 +40,20 @@ backends:
     type: vision_service
   vllm:
     api_key: ''
-    base_url: http://host.docker.internal:8000/v1
+    base_url: http://vllm:8000/v1
     enabled: true
     model: __auto__
     supports_images: false
     type: openai_compat
+    cf_orch:
+      service: vllm
+      model_candidates:
+      - Qwen2.5-3B-Instruct
+      - Phi-4-mini-instruct
+      ttl_s: 300
   vllm_research:
     api_key: ''
-    base_url: http://host.docker.internal:8000/v1
+    base_url: http://vllm:8000/v1
     enabled: true
     model: __auto__
     supports_images: false