feat(rag): 子章节级图片过滤 + VLM 后台增强 + 意图分析优化

- 图片选择新增 section_path 字段，支持子章节级过滤
- _filter_images_by_answer 增加 primary_sections 参数和叶子节点匹配
- 检索发散检测：primary_leaf_names > 3 时全局阈值+1
- _FIGURE_ANSWER_KEYWORDS 移除单字"图""表"，正则图号兜底防误触发
- lazy_enhance 后台增强流程优化
- 意图分析与 LLM 工具层改进

评测：图片选择 F1 从 52.4% 提升至 66.3%（+13.9pp），Precision +16.5pp

knowledge/collection.py

@@ -331,6 +331,21 @@ class CollectionMixin:
             except Exception as e:
                 logger.warning(f"清理版本记录失败: {e}")
 
+            # 清理不再被引用的图片和 VLM 缓存文件
+            # 注意：此时 ChromaDB collection 已删除，cleanup_image_orphans 会扫描
+            # 所有剩余 collection，仅该 collection 引用的图片会被识别为孤儿
+            try:
+                from knowledge.image_cleanup import cleanup_image_orphans
+                cleanup_result = cleanup_image_orphans(self)
+                if cleanup_result['deleted_images'] or cleanup_result['deleted_caches']:
+                    logger.info(
+                        f"清理孤儿文件: {cleanup_result['deleted_images']} 图片 + "
+                        f"{cleanup_result['deleted_caches']} VLM缓存, "
+                        f"释放 {cleanup_result['freed_bytes']/1024:.1f} KB"
+                    )
+            except Exception as e:
+                logger.warning(f"清理孤儿文件失败: {e}")
+
             if kb_name in self._metadata.get("collections", {}):
                 del self._metadata["collections"][kb_name]
             self._save_metadata()