feat: 工具调用全部异步，不再阻塞主会话

- 去掉8s超时同步执行，工具一律走 executeAsyncAndStore
- LLM立即收到[后台执行中]提示，继续生成回复
- 工具结果在下一轮对话中返回

Co-Authored-By: Claude <noreply@anthropic.com>

backend/ai-core/internal/orchestrator/synthesizer.go

@@ -87,32 +87,16 @@ func (s *Synthesizer) Synthesize(ctx context.Context, params SynthesizeParams, e
 				args = make(map[string]interface{})
 			}
 
+			s.emitToolProgress(eventCh, tc.Name, "started", 0, "正在执行 "+tc.Name)
 			s.emitToolProgress(eventCh, tc.Name, "started", 0, "正在执行 "+tc.Name)
 
-			toolCtx, cancel := context.WithTimeout(ctx, toolDeadline)
+			// 工具调用全部异步执行，不阻塞主会话
-			result, execErr := s.toolRegistry.Execute(toolCtx, tc.Name, args)
-			cancel()
-
-			if execErr != nil {
-				logger.Printf("[synthesizer] 工具 %s 执行失败: %v", tc.Name, execErr)
-			}
-			if result == nil {
-				result = &plgSDK.ToolResult{ToolName: tc.Name, Success: false, Error: execErr.Error()}
-			}
-
-			// Async fallback: if tool timed out, store for next turn
-			if toolCtx.Err() == context.DeadlineExceeded {
-				s.emitToolProgress(eventCh, tc.Name, "running", 0.5, tc.Name+" 执行时间较长，转入后台继续...")
 			go s.executeAsyncAndStore(tc, args, params.SessionID, eventCh)
-				result = &plgSDK.ToolResult{
+			result := &plgSDK.ToolResult{
 				ToolName: tc.Name,
 				Success:  true,
-					Output:   fmt.Sprintf("[后台执行中] %s 正在后台运行，结果将在下一轮对话中返回。你可以继续聊天。", tc.Name),
+				Output:   fmt.Sprintf("[后台执行中] %s 正在后台运行，结果稍后返回。", tc.Name),
 			}
-			} else {
-				s.emitToolProgress(eventCh, tc.Name, "completed", 1.0, "")
-			}
-
 			resultJSON, _ := json.Marshal(result)
 			messages = append(messages, model.LLMMessage{
 				Role:       model.RoleTool,