fix(auto_reply): 修复复合标题误匹配问题

- 方言语气词过滤：_STOPWORDS 新增"那搁""搁哪儿""哪儿"等东北话
- 复合标题检测：检测 heading 含 <br> 或列表标记时，只检查标题主体部分（<br>之前）的 meaningful tokens
- 防止 protect.md 的"温馨提示"body 含"登录"+"下载"被误选
- 修复'那搁哪儿下载那个登录插件啊?'返回错误文档的问题

auto_reply.py

@@ -89,6 +89,12 @@ class AutoReply:
         "给","把","被","让","请","该","此","为","等","其","将","时","后","前",
         "会","能","要","想","没","不","别","快","如","若","知","都","也","就",
         "已","再","更","最","很","太","较","但","却","因","故","则","且",
+        # 方言语气词（东北话等）——用户常问"那搁哪儿"、"咋整"等，这些不应参与核心词匹配
+        "那搁", "搁哪儿", "去哪儿", "哪儿", "哪能", "咋地", "咋整", "咋说",
+        # 疑问代词
+        "哪个", "那个", "这个", "哪些", "什么",
+        # 语气助词
+        "啊", "呀", "么", "嘛",
         # 通用名词，出现在太多文档中导致误匹配
         "服务器","玩家","游戏","世界","地图","技能","装备","物品","命令","指令",
     }
@@ -135,6 +141,19 @@ class AutoReply:
         has_ht = bool(q_mean & ht_tokens)
         has_body = bool(q_mean & body_tokens)
 
+        # ⭐ 复合标题检测（2026-08-13）：如果 heading 含 <br> 或列表标记，
+        # 说明是 Markdown 链接列表（如"- [登录]: url"），只检查标题主体部分（<br>之前）
+        # 的 meaningful tokens，而非整个复合 heading
+        # 防止 protect.md 的"温馨提示"body 含"登录"+"下载"被误选
+        _is_composite_heading = ('<br>' in heading or '- [' in heading or '|-' in heading)
+        if _is_composite_heading:
+            # 提取标题主体部分（<br>或列表之前的部分）
+            _title_part = re.split(r'<br>|- |\|-', heading_clean)[0]
+            _body_part_tokens = set(AutoReply._meaningful_tokens(_title_part))
+            # 重新计算 has_ht 为仅主体部分命中
+            has_ht = bool(q_mean & _body_part_tokens)
+            logger.debug(f"复合标题检测: heading={heading[:40]!r} title_part={_title_part!r} has_ht={has_ht}")
+
         # 动作/配置类查询：标题和正文都必须含核心词，防止"牛头不对马嘴"误匹配
         # ⭐ 扩展：has_body=False（正文完全无共同词）时直接拒绝，无论 heading 是否命中
         # 例："领地系统有哪些操作模式?" vs sdf1_login CLI段落 → 拒绝