sync(server-release): 从 main 同步逻辑改动（不含 API 格式变更）

同步内容:
- knowledge/manager.py: VLM max_tokens 512→2048 + reasoning_content fallback
- services/feedback.py: FAQ max_tokens 200→512
- services/session.py: 消息排序增加 id DESC 次级排序
- knowledge/image_cleanup.py: 新增图片/VLM缓存孤儿文件清理模块
- knowledge/collection.py: 集合删除时清理孤儿文件 + list_collections 磁盘扫描策略
- knowledge/document.py: 文档删除时清理孤儿文件
- api/chat_routes.py: 新增 _rescue_bm25_divergence 函数（BM25-CE分歧救援）
- core/intent_analyzer.py: 使用 get_intent_client 专用客户端 + 移除短追问缓存跳过逻辑
- cleanup_orphans.py: 独立孤儿文件清理脚本

knowledge/image_cleanup.py

@@ -0,0 +1,140 @@
+"""
+图片/VLM缓存孤儿文件清理模块
+
+提供可被 document.py / collection.py 调用的清理函数，
+也可被 cleanup_orphans.py 独立脚本使用。
+"""
+
+import hashlib
+import logging
+import os
+from pathlib import Path
+
+logger = logging.getLogger(__name__)
+
+IMAGES_DIR = Path(".data/images")
+VLM_CACHE_DIR = Path(".data/cache/vlm")
+
+
+def compute_file_hash(file_path: str) -> str:
+    """计算文件 MD5"""
+    with open(file_path, 'rb') as f:
+        return hashlib.md5(f.read()).hexdigest()
+
+
+def collect_referenced_images(manager, collections=None) -> set:
+    """
+    从 ChromaDB 收集所有被引用的图片文件名。
+
+    Args:
+        manager: KnowledgeBaseManager 实例
+        collections: 限定知识库列表，None 表示全部
+
+    Returns:
+        set of image filenames (e.g., {"185a7a75d246.png", ...})
+    """
+    referenced = set()
+
+    if collections:
+        kb_names = collections
+    else:
+        kb_names = [c.name if hasattr(c, 'name') else str(c)
+                     for c in manager.list_collections()]
+
+    for kb_name in kb_names:
+        try:
+            col = manager.get_collection(kb_name)
+        except Exception:
+            continue
+
+        results = col.get(include=['metadatas'])
+        if not results['ids']:
+            continue
+
+        for meta in results['metadatas']:
+            image_path = meta.get('image_path', '')
+            if image_path:
+                referenced.add(os.path.basename(image_path))
+
+    return referenced
+
+
+def cleanup_image_orphans(manager, collections=None, dry_run=False) -> dict:
+    """
+    清理不再被任何 ChromaDB 切片引用的图片和 VLM 缓存文件。
+
+    Args:
+        manager: KnowledgeBaseManager 实例
+        collections: 限定知识库列表，None 表示全部
+        dry_run: True 时只返回孤儿列表不实际删除
+
+    Returns:
+        {
+            'orphan_images': [(filepath, filename, size_bytes)],
+            'orphan_caches': [(filepath, filename, size_bytes)],
+            'deleted_images': int,
+            'deleted_caches': int,
+            'freed_bytes': int
+        }
+    """
+    result = {
+        'orphan_images': [],
+        'orphan_caches': [],
+        'deleted_images': 0,
+        'deleted_caches': 0,
+        'freed_bytes': 0
+    }
+
+    # 1. 收集引用
+    referenced = collect_referenced_images(manager, collections)
+
+    # 2. 查找孤儿图片
+    if IMAGES_DIR.exists():
+        for f in IMAGES_DIR.iterdir():
+            if f.is_file() and f.name not in referenced:
+                result['orphan_images'].append((str(f), f.name, f.stat().st_size))
+
+    # 3. 查找孤儿 VLM 缓存（图片已删除则缓存也应是孤儿）
+    if VLM_CACHE_DIR.exists():
+        referenced_hashes = set()
+        for filename in referenced:
+            full_path = IMAGES_DIR / filename
+            if full_path.exists():
+                try:
+                    img_hash = compute_file_hash(str(full_path))
+                    referenced_hashes.add(img_hash)
+                except Exception:
+                    pass
+
+        for f in VLM_CACHE_DIR.iterdir():
+            if f.is_file() and f.suffix == '.txt':
+                cache_hash = f.stem
+                if cache_hash not in referenced_hashes:
+                    result['orphan_caches'].append((str(f), f.name, f.stat().st_size))
+
+    # 4. 删除
+    if not dry_run:
+        for filepath, filename, size in result['orphan_images']:
+            try:
+                os.remove(filepath)
+                result['deleted_images'] += 1
+                result['freed_bytes'] += size
+            except OSError as e:
+                logger.warning(f"删除孤儿图片失败: {filename} - {e}")
+
+        for filepath, filename, size in result['orphan_caches']:
+            try:
+                os.remove(filepath)
+                result['deleted_caches'] += 1
+                result['freed_bytes'] += size
+            except OSError as e:
+                logger.warning(f"删除孤儿缓存失败: {filename} - {e}")
+
+        if result['deleted_images'] or result['deleted_caches']:
+            logger.info(
+                f"清理孤儿: {result['deleted_images']} 图片 + "
+                f"{result['deleted_caches']} 缓存, "
+                f"释放 {result['freed_bytes']/1024:.1f} KB"
+            )
+
+    return result