feat: expose graph orchestration mode, fix pipeline multi-agent, add Feishu tools (Phase 3)

增强编排 + 飞书深度集成：
- Graph 模式：暴露 orchestrator._graph() 到 run() 方法，workflow_integration 支持 graph nodes/edges
- Pipeline 修复：多 Agent 按步骤轮转分配，不再只用 agents[0]
- 4个飞书操作工具: feishu_create_doc / feishu_create_calendar_event / feishu_search_contacts / feishu_send_approval
- 飞书 @mention→Goal：feishu/ orange WS handler 支持 "目标: xxx" 触发自动创建 Goal

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

backend/app/agent_runtime/orchestrator.py

@@ -151,8 +151,19 @@ class AgentOrchestrator:
         question: str,
         agents: List[OrchestratorAgentConfig],
         on_llm_call: Optional[Callable[[Dict[str, Any]], Any]] = None,
+        graph_nodes: Optional[List[Dict[str, Any]]] = None,
+        graph_edges: Optional[List[Dict[str, Any]]] = None,
     ) -> OrchestratorResult:
-        """执行多 Agent 编排。"""
+        """执行多 Agent 编排。
+
+        Args:
+            mode: route / sequential / debate / pipeline / graph
+            question: 用户问题
+            agents: Agent 配置列表
+            on_llm_call: LLM 调用回调
+            graph_nodes: graph 模式的节点定义（mode=graph 时必填）
+            graph_edges: graph 模式的边定义（mode=graph 时必填）
+        """
         mode = mode.lower()
         if mode == "route":
             return await self._route(question, agents, on_llm_call)
@@ -162,8 +173,12 @@ class AgentOrchestrator:
             return await self._debate(question, agents, on_llm_call)
         elif mode == "pipeline":
             return await self._pipeline(question, agents, on_llm_call)
+        elif mode == "graph":
+            if not graph_nodes:
+                raise ValueError("graph 模式需要提供 graph_nodes 参数")
+            return await self._graph(question, graph_nodes, graph_edges or [], on_llm_call)
         else:
-            raise ValueError(f"不支持的编排模式: {mode}，可选: route, sequential, debate, pipeline")
+            raise ValueError(f"不支持的编排模式: {mode}，可选: route, sequential, debate, pipeline, graph")
 
     async def _route(
         self, question: str, agents: List[OrchestratorAgentConfig],
@@ -500,11 +515,13 @@ class AgentOrchestrator:
                 steps=steps,
             )
 
-        # ── 2. Executor：逐步骤执行 ──
-        executor_cfg = agents[0] if agents else OrchestratorAgentConfig(
-            id="executor", name="Executor",
-            system_prompt="你是一个有用的AI助手。",
-        )
+        # ── 2. Executor：逐步骤执行（多 Agent 轮转分配）──
+        executor_pool = agents if agents else [
+            OrchestratorAgentConfig(
+                id="executor", name="Executor",
+                system_prompt="你是一个有用的AI助手。",
+            )
+        ]
 
         previous_output = "（尚无前序步骤）"
         execution_results = []
@@ -514,6 +531,9 @@ class AgentOrchestrator:
             step_desc = step_info.get("description", f"步骤 {step_num}")
             step_expect = step_info.get("expected_output", "")
 
+            # 按步骤轮转分配 Agent：不同步骤可分配给不同 Agent（按专长匹配）
+            executor_cfg = executor_pool[(step_num - 1) % len(executor_pool)]
+
             executor_prompt = _EXECUTOR_STEP_PROMPT.format(
                 original_question=question,
                 plan_title=plan.get("plan_title", ""),
@@ -555,6 +575,7 @@ class AgentOrchestrator:
             execution_results.append({
                 "step": step_num,
                 "description": step_desc,
+                "agent": executor_cfg.name,
                 "output": step_result.content,
                 "error": step_result.error if not step_result.success else None,
             })
@@ -562,7 +583,7 @@ class AgentOrchestrator:
             previous_output = step_result.content if step_result.success else f"（步骤{step_num}执行出错）"
 
             if not step_result.success:
-                logger.warning(f"Pipeline 步骤{step_num} 执行失败: {step_result.error}")
+                logger.warning(f"Pipeline 步骤{step_num} ({executor_cfg.name}) 执行失败: {step_result.error}")
 
         # ── 3. Reviewer：审查并交付 ──
         plan_steps_text = "\n".join(