feat: 平台规则从硬编码改为品牌方上传文档 + AI 解析

- 新增 PlatformRule 模型 (draft/active/inactive 状态流转)
- 新增文档解析服务 (PDF/Word/Excel → 纯文本)
- 新增 4 个 API: 解析/确认/查询/删除平台规则
- 脚本审核优先从 DB 读取 active 规则，硬编码兜底
- 视频审核合并平台规则违禁词到检测列表
- Alembic 迁移 006: platform_rules 表

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

backend/app/tasks/review.py

@@ -14,7 +14,7 @@ from sqlalchemy.orm import sessionmaker
 
 from app.config import settings
 from app.models.review import ReviewTask, TaskStatus as DBTaskStatus
-from app.models.rule import ForbiddenWord, Competitor
+from app.models.rule import ForbiddenWord, Competitor, PlatformRule, RuleStatus
 from app.models.ai_config import AIConfig
 from app.services.video_download import VideoDownloadService, DownloadResult
 from app.services.keyframe import KeyFrameExtractor, ExtractionResult
@@ -81,6 +81,7 @@ async def complete_review(
     summary: str,
     violations: list[dict],
     status: DBTaskStatus = DBTaskStatus.COMPLETED,
+    soft_warnings: Optional[list[dict]] = None,
 ):
     """完成审核"""
     result = await db.execute(
@@ -94,6 +95,8 @@ async def complete_review(
         task.score = score
         task.summary = summary
         task.violations = violations
+        if soft_warnings is not None:
+            task.soft_warnings = soft_warnings
         task.completed_at = datetime.now(timezone.utc)
         await db.commit()
 
@@ -153,6 +156,24 @@ async def get_competitors(db: AsyncSession, tenant_id: str, brand_id: str) -> li
     return [row[0] for row in result.fetchall()]
 
 
+async def get_platform_forbidden_words(
+    db: AsyncSession, tenant_id: str, brand_id: str, platform: str,
+) -> list[str]:
+    """从 DB 获取品牌方在该平台的 active 规则中的违禁词"""
+    result = await db.execute(
+        select(PlatformRule).where(
+            PlatformRule.tenant_id == tenant_id,
+            PlatformRule.brand_id == brand_id,
+            PlatformRule.platform == platform,
+            PlatformRule.status == RuleStatus.ACTIVE.value,
+        )
+    )
+    rule = result.scalar_one_or_none()
+    if not rule or not rule.parsed_rules:
+        return []
+    return rule.parsed_rules.get("forbidden_words", [])
+
+
 async def process_video_review(
     review_id: str,
     tenant_id: str,
@@ -199,6 +220,13 @@ async def process_video_review(
 
             # 获取规则
             forbidden_words = await get_forbidden_words(db, tenant_id)
+            # 合并平台规则中的违禁词
+            platform_fw = await get_platform_forbidden_words(db, tenant_id, brand_id, platform)
+            existing_set = set(forbidden_words)
+            for w in platform_fw:
+                if w not in existing_set:
+                    forbidden_words.append(w)
+                    existing_set.add(w)
             competitors = await get_competitors(db, tenant_id, brand_id)
 
             # 初始化 AI 服务
@@ -281,16 +309,37 @@ async def process_video_review(
                 )
                 all_violations.extend(subtitle_violations)
 
-            # 6. 计算分数和生成报告
+            # 6. 分流 violations / soft_warnings
             await update_review_progress(db, review_id, 90, "生成报告")
-            score = review_service.calculate_score(all_violations)
 
-            if not all_violations:
+            hard_violations = []
+            soft_warnings_data = []
+
+            for v in all_violations:
+                v_type = v.get("type", "")
+                if v_type in ("forbidden_word", "efficacy_claim", "competitor_logo", "brand_safety"):
+                    hard_violations.append(v)
+                elif v_type in ("duration_short", "mention_missing"):
+                    soft_warnings_data.append({
+                        "code": f"video_{v_type}",
+                        "message": v.get("content", ""),
+                        "action_required": "note",
+                        "blocking": False,
+                        "context": {"suggestion": v.get("suggestion", "")},
+                    })
+                else:
+                    hard_violations.append(v)  # 默认当硬性违规
+
+            # 计算分数（仅硬性违规影响分数）
+            score = review_service.calculate_score(hard_violations)
+
+            if not hard_violations:
                 summary = "视频内容合规，未发现违规项"
+                if soft_warnings_data:
+                    summary += f"（{len(soft_warnings_data)} 条提醒）"
             else:
-                high_count = sum(1 for v in all_violations if v.get("risk_level") == "high")
-                medium_count = sum(1 for v in all_violations if v.get("risk_level") == "medium")
-                summary = f"发现 {len(all_violations)} 处违规"
+                high_count = sum(1 for v in hard_violations if v.get("risk_level") == "high")
+                summary = f"发现 {len(hard_violations)} 处违规"
                 if high_count > 0:
                     summary += f"（{high_count} 处高风险）"
 
@@ -300,7 +349,8 @@ async def process_video_review(
                 review_id,
                 score=score,
                 summary=summary,
-                violations=all_violations,
+                violations=hard_violations,
+                soft_warnings=soft_warnings_data if soft_warnings_data else None,
             )
 
         except Exception as e: