feat(脚本): 按流程文档重写脚本阶段(自动提取人物/场景 + 标签胶囊改写 + 基础资产按标签 seed)

后端:
- 出稿后新增轻量提取调用,从脚本抽取人物/场景标签及每个标签的建议生图提示词,
  写入 project.metadata(cast/scenes/cast_prompts/scene_prompts)
- 提取调用接入计费闭环(reserve→charge/release,记为 script_optimization);
  best-effort:无模型/预扣失败/调用失败/解析失败都吞掉返回空,绝不阻断出稿
- generate_base_asset 增加 label,落 group.metadata,把生成结果归回脚本标签卡

前端:
- 设定卡风格补齐到 5 个(真实测评/痛点种草/小红书种草/开箱测评/对比展示)
- 标签改为待确认胶囊流:增删不立即落库,排进队列在输入框上方显示可撤销胶囊,
  点发送才整体重写;待删划线、待加虚线
- 基础资产人物/场景按提取标签逐个 seed 卡,显示 AI 提示词、可改后生成

测试:新增提取+计费+release+label 共 4 个用例,全套 24 个通过(sqlite test 设置)

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

core/backend/apps/projects/views.py

@@ -157,7 +157,7 @@ class ProjectViewSet(TeamScopedViewSetMixin, ModelViewSet):
         kind = request.data.get("kind")
         if kind not in BaseAssetGroup.Kind.values:
             return Response({"detail": "invalid base asset kind"}, status=status.HTTP_400_BAD_REQUEST)
-        group = generate_base_asset(project=project, user=request.user, kind=kind, prompt=request.data.get("prompt", ""))
+        group = generate_base_asset(project=project, user=request.user, kind=kind, prompt=request.data.get("prompt", ""), label=request.data.get("label", ""))
         stage, _ = ProjectStage.objects.get_or_create(project=project, stage=ProjectStage.Stage.BASE_ASSETS)
         stage.status = ProjectStage.Status.NEEDS_REVIEW
         stage.save(update_fields=["status", "updated_at"])