fix: 优化同轮语义去重——降低关键词重叠阈值至25%+动作词检测

- 骨架完全相同 → 跳过
- SequenceMatcher相似度>=70% → 跳过
- 关键词重叠>=25% 或 含核心动作词且重叠>=3 → 跳过

修复场景："我想知道那个怎么撤销导入的兑换码" vs "那啥怎么去撤销那个导入过的激活码"
两消息共享动作词(撤销/导入)和4个关键词，判定为重复

monitor.py

@@ -10184,22 +10184,38 @@ class MessageMonitor:
                         # 使用骨架匹配快速判断
                         sk_cur = _cn_skeleton(_cleaned)
                         sk_prev = _cn_skeleton(_prev_cleaned)
-                        if len(sk_cur) >= 6 and sk_prev and sk_cur == sk_prev:
-                            logger.info(
-                                f"[{platform}] 同轮语义去重: 骨架相同，跳过: {_cleaned[:40]!r} ~ {_prev_cleaned[:40]!r}"
-                            )
-                            _is_same_round_dup = True
-                            break
-                        # Jaccard相似度兜底
-                        if len(sk_cur) >= 4 and len(sk_prev) >= 4:
+                        if len(sk_cur) >= 6 and sk_prev:
+                            # 精确匹配
+                            if sk_cur == sk_prev:
+                                logger.info(
+                                    f"[{platform}] 同轮语义去重: 骨架相同，跳过: {_cleaned[:40]!r} ~ {_prev_cleaned[:40]!r}"
+                                )
+                                _is_same_round_dup = True
+                                break
+                            # 高相似度匹配
                             from difflib import SequenceMatcher
                             sim = SequenceMatcher(None, sk_cur, sk_prev).ratio()
-                            if sim >= 0.85:
+                            if sim >= 0.7:
                                 logger.info(
                                     f"[{platform}] 同轮语义去重: 相似度{sim:.0%}，跳过: {_cleaned[:40]!r} ~ {_prev_cleaned[:40]!r}"
                                 )
                                 _is_same_round_dup = True
                                 break
+                            # 关键词重叠检测（针对长文本）
+                            if len(sk_cur) >= 8 and len(sk_prev) >= 8:
+                                kw_cur = {sk_cur[i:i+2] for i in range(len(sk_cur)-1)}
+                                kw_prev = {sk_prev[i:i+2] for i in range(len(sk_prev)-1)}
+                                overlap = kw_cur & kw_prev
+                                total = kw_cur | kw_prev
+                                # 重叠度>=25%或包含核心动作词则判定为重复
+                                _action_kw = {'撤销', '导入', '兑换', '激活', '查询', '打印', '操作'}
+                                has_action = bool(kw_cur & _action_kw) and bool(kw_prev & _action_kw)
+                                if total and (len(overlap) / len(total) >= 0.25 or (has_action and len(overlap) >= 3)):
+                                    logger.info(
+                                        f"[{platform}] 同轮语义去重: 关键词重叠{len(overlap)}/{len(total)}，跳过: {_cleaned[:40]!r} ~ {_prev_cleaned[:40]!r}"
+                                    )
+                                    _is_same_round_dup = True
+                                    break
                     if _is_same_round_dup:
                         continue