fix(extract): 实体提取改异步(worker+轮询),根治 502 + 刷新丢 loading + 重复扣费

问题:进资产页点"提取人物/场景"——同步死等豆包思考模型(数十秒)→ 撞网关超时
502;"提取中"loading 是纯前端内存态,刷新即丢;且无幂等,刷新后重点会重复扣费。

改动(对齐全站异步三件套:出图/三视图/故事板同款):
- 后端 submit_extract_entities:Web 只建 RESERVED 任务 + 预留额度秒回(不再 502);
  已有在途提取则复用,绝不二次预扣(防刷新后重点 / 并发重复扣费)。
- 后端 run_extract_entities_task(Celery worker):流式调豆包 + 解析 + 落库 + 扣费;
  失败退费并把可读错误记进 task.error_message。
- 新增 AITask.Type.entity_extraction(迁移 0011)+ extract-status 端点(轮询进度/成败)。
- 前端 runExtract 改提交+轮询;进资产页自动认领在途提取,刷新后 loading 自己回来。

验收:完整 Django 套件 149/149 全绿(含新增 4 测:异步落库+计费一次 / content空
回退 reasoning / 在途防重复扣费 / 状态端点成败透出);tsc + vite build 全绿;
makemigrations --check 无遗漏。

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

core/backend/apps/projects/views.py

@@ -19,13 +19,14 @@ from apps.ai.services import (
     DEFAULT_VOICEOVER_VOICE,
     VOICEOVER_VOICES,
     create_export_job,
-    extract_entities_for_project,
     generate_base_asset,
     generate_person_triview,
     generate_storyboard_frame,
     get_default_model,
+    get_inflight_extraction,
     poll_video_segment,
     regenerate_script_segment,
+    submit_extract_entities,
     submit_storyboard,
     submit_video_segment,
     synthesize_project_voiceover,
@@ -294,15 +295,37 @@ class ProjectViewSet(TeamScopedViewSetMixin, ModelViewSet):
 
     @action(detail=True, methods=["post"], url_path="extract-entities")
     def extract_entities_action(self, request, pk=None):
-        """从已定稿脚本提取角色 / 场景实体(独立步,进资产阶段时调)。AI 读剧本 → 角色 / 场景 + 每镜 refs,
-        落库覆盖 project.metadata + 每镜 entity_refs;**商品不提**(参考图层用真实主图 / 三视图无条件带上)。
-        前端进资产页蒙版上「只提取 / 提取+生成」按钮点的就是它(生成由前端按返回实体循环调 generate-base-asset)。"""
+        """从已定稿脚本提取角色 / 场景实体(独立步,进资产阶段时调)——**异步**:豆包思考模型流式抽取较慢,
+        Web 层只建任务秒回(不再 502),慢活交 worker(run_extract_entities_task)。前端拿 task_id 轮询
+        extract-status 取结果。已有在途提取则复用(防刷新后重点重复扣费)。落库覆盖 project.metadata +
+        每镜 entity_refs;**商品不提**(参考图层用真实主图)。"""
+        require_worker()  # 异步提取依赖 worker 兜底执行,没 worker 直接拒绝(否则任务永远 RESERVED)
         project = self.get_object()
         try:
-            result = extract_entities_for_project(project=project, user=request.user)
-        except ValueError as exc:  # 无脚本 / 无模型 / 额度不足 / 解析失败 等,立即反馈
+            task = submit_extract_entities(project=project, user=request.user)
+        except ValueError as exc:  # 无脚本 / 无分镜 / 无模型 / 额度不足,立即反馈
             return Response({"detail": str(exc)}, status=status.HTTP_400_BAD_REQUEST)
-        return Response(result)
+        return Response({"task_id": str(task.id), "status": task.status})
+
+    @action(detail=True, methods=["get"], url_path="extract-status")
+    def extract_status_action(self, request, pk=None):
+        """实体提取进度(异步)。前端进资产页 / 点提取后据此轮询 + 刷新后重建「提取中」loading:
+        running=有在途任务;否则读最近一条提取任务的成败。成功时回带 entities(供「提取+生成」继续出参考图)。"""
+        project = self.get_object()
+        inflight = get_inflight_extraction(project)
+        if inflight is not None:
+            return Response({"running": True, "status": inflight.status, "error": "", "entities": []})
+        last = (
+            AITask.objects.filter(project=project, task_type=AITask.Type.ENTITY_EXTRACTION)
+            .order_by("-created_at")
+            .first()
+        )
+        if last is None:
+            return Response({"running": False, "status": "", "error": "", "entities": []})
+        entities = (project.metadata or {}).get("script_entities", []) if last.status == AITask.Status.SUCCEEDED else []
+        return Response(
+            {"running": False, "status": last.status, "error": last.error_message or "", "entities": entities}
+        )
 
     @action(detail=True, methods=["post"], url_path="generate-base-asset")
     def generate_base_asset_action(self, request, pk=None):