feat: 完善差旅票据行程提取与费用明细回填逻辑

增强文档智能识别的票据场景关键词和字段提取能力，优化
会话关联草稿报销单的解析路径，修复费用明细合并和票据
去重边界问题，前端改进报销创建和审批详情交互，补充单
元测试覆盖。

server/src/app/services/agent_conversations.py

@@ -18,6 +18,47 @@ STATEFUL_CONTEXT_KEYS = (
     "attachment_count",
     "ocr_summary",
     "ocr_documents",
+    "review_form_values",
+    "business_time_context",
+)
+REVIEW_FLOW_CONTEXT_KEYS = {
+    "request_context",
+    "attachment_names",
+    "attachment_count",
+    "ocr_summary",
+    "ocr_documents",
+    "review_form_values",
+    "business_time_context",
+}
+REVIEW_FLOW_CONTINUATION_KEYWORDS = (
+    "补充",
+    "继续",
+    "继续上传",
+    "当前",
+    "这张",
+    "这个",
+    "该单据",
+    "现有",
+    "已有",
+    "关联",
+    "合并",
+    "修改",
+    "更正",
+    "改成",
+    "调整",
+    "下一步",
+    "保存草稿",
+)
+NEW_EXPENSE_PROMPT_KEYWORDS = (
+    "申请报销",
+    "我要报销",
+    "我想报销",
+    "帮我报销",
+    "发起报销",
+    "提交报销",
+    "生成报销",
+    "创建报销",
+    "新建报销",
 )
 DEFAULT_CONVERSATION_RETENTION_DAYS = 3
 
@@ -182,10 +223,15 @@ class AgentConversationService:
         *,
         conversation: AgentConversation,
         context_json: dict[str, Any],
+        message: str | None = None,
         history_limit: int = 8,
     ) -> dict[str, Any]:
         merged = dict(context_json or {})
         state_json = dict(conversation.state_json or {})
+        should_hydrate_review_flow = self._should_hydrate_review_flow_context(
+            context_json=merged,
+            message=message,
+        )
 
         merged["conversation_id"] = conversation.conversation_id
         merged["conversation_history"] = self.list_message_history(
@@ -196,16 +242,53 @@ class AgentConversationService:
             merged.setdefault("conversation_scenario", conversation.last_scenario)
         if conversation.last_intent:
             merged.setdefault("conversation_intent", conversation.last_intent)
-        if conversation.draft_claim_id and not str(merged.get("draft_claim_id") or "").strip():
+        if (
+            should_hydrate_review_flow
+            and conversation.draft_claim_id
+            and not str(merged.get("draft_claim_id") or "").strip()
+        ):
             merged["draft_claim_id"] = conversation.draft_claim_id
         merged["conversation_state"] = state_json
 
         for key in STATEFUL_CONTEXT_KEYS:
+            if key in REVIEW_FLOW_CONTEXT_KEYS and not should_hydrate_review_flow:
+                continue
             if self._is_empty_value(merged.get(key)) and not self._is_empty_value(state_json.get(key)):
                 merged[key] = state_json.get(key)
 
         return merged
 
+    @staticmethod
+    def _should_hydrate_review_flow_context(
+        *,
+        context_json: dict[str, Any],
+        message: str | None,
+    ) -> bool:
+        if AgentConversationService._resolve_draft_claim_id(context_json):
+            return True
+        if str(context_json.get("review_action") or "").strip():
+            return True
+        if str(context_json.get("entry_source") or "").strip() == "detail":
+            return True
+        if not AgentConversationService._is_empty_value(context_json.get("attachment_names")):
+            return True
+        if not AgentConversationService._is_empty_value(context_json.get("ocr_documents")):
+            return True
+        if str(context_json.get("ocr_summary") or "").strip():
+            return True
+        try:
+            if int(context_json.get("attachment_count") or 0) > 0:
+                return True
+        except (TypeError, ValueError):
+            pass
+
+        compact_message = str(message or "").replace(" ", "")
+        if not compact_message:
+            return False
+        if any(keyword in compact_message for keyword in NEW_EXPENSE_PROMPT_KEYWORDS):
+            return False
+        return any(keyword in compact_message for keyword in REVIEW_FLOW_CONTINUATION_KEYWORDS)
+
     def append_message(
         self,
         *,