refactor(server): steward 决策链路改用 LangGraph 编排

- 新增 StewardGraphPlannerService，用 LangGraph 状态图编排意图识别→流程判断→模型/规则分支→兜底，替代原 planner 内线性调用
- 新增 StewardGraphRuntimeService 编排运行时决策与槽位决策；StewardActionContracts/Executor 统一动作合约与执行
- steward_intent_agent/application_fact_resolver/runtime_chat 适配图执行器，config 暴露图相关开关
- pyproject/uv.lock 新增 langgraph 依赖
- 新增 graph_planner/graph_runtime/action_executor 测试，更新 intent_agent/planner/fact_resolver/runtime_chat/reimbursement 测试

server/src/app/services/steward_planner_fallback.py

@@ -34,9 +34,8 @@ class StewardPlannerFallbackMixin:
         base_date: date,
         request: StewardPlanRequest,
     ) -> bool:
-        if self._looks_like_ambiguous_travel_flow(message, base_date, request):
-            return False
-        return self._has_multiple_financial_demands(message)
+        return bool(message.strip())
+
 
     @staticmethod
     def _is_business_irrelevant_input(message: str, request: StewardPlanRequest) -> bool:
@@ -105,12 +104,14 @@ class StewardPlannerFallbackMixin:
         request: StewardPlanRequest,
         *,
         scenario: str,
+        model_call_traces: list[dict[str, Any]] | None = None,
+        fallback_reason: str = "",
     ) -> StewardPlanResponse:
         """业务无关输入的兜底计划：根据场景给出对应引导，off_business 场景可由 LLM 增强。"""
         base_summary = self._default_off_topic_summary(scenario)
         thinking_event = self._build_off_topic_thinking_event(scenario)
         suggested_prompts = self._off_topic_suggested_prompts(scenario)
-        model_call_traces: list[dict[str, Any]] = []
+        traces = list(model_call_traces or [])
 
         # 仅对 off_business 场景尝试让 LLM 生成多样化引导；问候/无意义场景用规则模板即可。
         if (
@@ -121,24 +122,36 @@ class StewardPlannerFallbackMixin:
                 llm_result = self.off_topic_agent.generate(request, scenario=scenario)
                 if llm_result is not None and llm_result.response_text:
                     base_summary = llm_result.response_text
-                    model_call_traces = llm_result.model_call_traces
+                    traces = llm_result.model_call_traces
             except Exception:
                 # 失败时静默回退到规则模板
                 pass
 
+        thinking_events = [thinking_event]
+        if fallback_reason:
+            thinking_events.insert(
+                0,
+                StewardThinkingEvent(
+                    event_id="intent_agent_rule_fallback",
+                    stage="rule_fallback",
+                    title="意图识别智能体进入兜底模式",
+                    content=fallback_reason,
+                ),
+            )
+
         return StewardPlanResponse(
             plan_id=f"steward_plan_{uuid.uuid4().hex[:12]}",
             plan_status="off_topic",
             planning_source="rule_fallback",
             next_action="none",
             summary=base_summary,
-            thinking_events=[thinking_event],
+            thinking_events=thinking_events,
             tasks=[],
             attachment_groups=[],
             confirmation_groups=[],
             candidate_flows=[],
             suggested_prompts=suggested_prompts,
-            model_call_traces=model_call_traces,
+            model_call_traces=traces,
         )
 
     @staticmethod
@@ -435,4 +448,3 @@ class StewardPlannerFallbackMixin:
             )
 
         return drafts
-