fix: 彻底修复 RAG 无限循环问题

关键修复：
1. rag_nodes.py: history 里记录的 action 改成大写 RETRIEVE_RAG
2. react_nodes.py: route_by_reasoning 优先检查 DIRECT_RESPONSE
3. 优化路由判断顺序，避免不必要的循环

backend/app/main_graph/nodes/rag_nodes.py

@@ -234,7 +234,7 @@ async def rag_retrieve_node(state: MainGraphState, config: Optional[Dict[str, An
             # 关键修复：把 rag_retrieve 加到 reasoning_history 里，让下次推理知道
             state.reasoning_history.append({
                 "step": state.reasoning_step,
-                "action": "rag_retrieve",
+                "action": "RETRIEVE_RAG",  # 大写，和推理结果保持一致
                 "confidence": 1.0,
                 "reasoning": "RAG 检索完成",
                 "timestamp": datetime.now().isoformat()