feat: 群聊并发 — 主会话繁忙时创建协会议话

- Orchestrator 新增 per-session 处理锁 + co-session 跟踪
- 主会话处理中时，新消息自动创建协会议话（并发 LLM 调用）
- 协会议话上限 3 个，超出排队等待 500ms
- 主会话结束后自动释放锁

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

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

@@ -43,6 +43,12 @@ type Orchestrator struct {
 	ocrProvider     llm.LLMProvider // OCR 模型 (文字提取，与视觉模型并行调用)
 	videoProvider   llm.LLMProvider // 视频模型 (短视频理解)
 	asrProvider     llm.ASRProvider  // ASR 语音识别 (语音消息转录)
+
+	// 群聊并发：主会话繁忙时创建协会议话
+	sessionProcMu   sync.Mutex
+	sessionProc     map[string]bool // sessionID → currently processing
+	activeCoSessions map[string]int // sessionID → active co-session count
+	maxCoSessions   int             // max concurrent co-sessions per main session (default 3)
 }
 
 // SetResponseCache sets the response cache (optional, for Phase 0.2).
@@ -116,14 +122,17 @@ func NewOrchestrator(
 	memoryExtractor *memory.Extractor,
 ) *Orchestrator {
 	return &Orchestrator{
-		personaLoader:   personaLoader,
-		contextBuilder:  contextBuilder,
-		llmAdapter:      chatAdapter,
-		subManager:      subManager,
-		intentAnalyzer:  NewIntentAnalyzer(intentAdapter),
-		synthesizer:     NewSynthesizer(chatAdapter, nil),
-		memoryRetriever: memoryRetriever,
-		memoryExtractor: memoryExtractor,
+		personaLoader:     personaLoader,
+		contextBuilder:    contextBuilder,
+		llmAdapter:        chatAdapter,
+		subManager:        subManager,
+		intentAnalyzer:    NewIntentAnalyzer(intentAdapter),
+		synthesizer:       NewSynthesizer(chatAdapter, nil),
+		memoryRetriever:   memoryRetriever,
+		memoryExtractor:   memoryExtractor,
+		sessionProc:       make(map[string]bool),
+		activeCoSessions:  make(map[string]int),
+		maxCoSessions:     3,
 	}
 }
 
@@ -172,7 +181,42 @@ func (o *Orchestrator) ProcessInput(
 			}
 		}()
 
-		// 0. 发布合成开始事件
+		// 0. 群聊并发：检测主会话是否繁忙，决定是主会话还是协会议话
+		isCoSession := false
+		o.sessionProcMu.Lock()
+		if o.sessionProc[params.SessionID] {
+			if o.activeCoSessions[params.SessionID] >= o.maxCoSessions {
+				o.sessionProcMu.Unlock()
+				logger.Printf("[orchestrator] 协会议话已达上限，排队等待")
+				time.Sleep(500 * time.Millisecond)
+				o.sessionProcMu.Lock()
+			}
+			o.activeCoSessions[params.SessionID]++
+			isCoSession = true
+			o.sessionProcMu.Unlock()
+			logger.Printf("[orchestrator] 主会话繁忙，创建协会议话 (session=%s, active=%d)", params.SessionID, o.activeCoSessions[params.SessionID])
+		} else {
+			o.sessionProc[params.SessionID] = true
+			o.sessionProcMu.Unlock()
+			defer func() {
+				o.sessionProcMu.Lock()
+				delete(o.sessionProc, params.SessionID)
+				o.sessionProcMu.Unlock()
+			}()
+		}
+
+		if isCoSession {
+			defer func() {
+				o.sessionProcMu.Lock()
+				o.activeCoSessions[params.SessionID]--
+				if o.activeCoSessions[params.SessionID] <= 0 {
+					delete(o.activeCoSessions, params.SessionID)
+				}
+				o.sessionProcMu.Unlock()
+			}()
+		}
+
+		// 0.5 发布合成开始事件
 		o.getBus().Publish(bus.BusEvent{
 			Type:      bus.EventSynthesisStarted,
 			SessionID: params.SessionID,