diff --git a/fam-edge/src/fam_edge/oracle_db.py b/fam-edge/src/fam_edge/oracle_db.py
index e1f09ef..2b87f8c 100644
--- a/fam-edge/src/fam_edge/oracle_db.py
+++ b/fam-edge/src/fam_edge/oracle_db.py
@@ -475,7 +475,35 @@ class OracleDB:
         # 带备注的原始输出分裂出垃圾人物行
         label = re.sub(r'[（(][^()（）]*[)）]', '', str(label)).strip() or str(label)
         now = _now_iso()
+        display_uid = display_uid or label
         row = self._conn.execute("SELECT * FROM people WHERE label=?", (label,)).fetchone()
+        # 跨视频 uid 复用检测：大模型给的 人物A/B/C 这类 uid 只在单次视频分析内部
+        # 稳定，不同视频各自独立编号——同一个字符串在不同视频里完全可能指向不同的
+        # 真人（实测 "人物A" 在 29 个视频里混了至少男女两个人）。这里用性别这个相对
+        # 稳定的信号做冲突检测：这次新特征和已有行的性别都明确且互相矛盾，说明大概
+        # 率撞车到了另一个人，不能直接合并覆盖——改挂到一个新的派生 label（如
+        # "人物A#2"）上，留给 person_service 的特征合并再判断这个新 label 该并进
+        # 哪个已命名身份，避免把两个人的特征越merge越乱。
+        if row and features:
+            try:
+                old_gender = (json.loads(row['features_json'] or '{}') or {}).get('gender', '')
+            except (ValueError, TypeError):
+                old_gender = ''
+            old_gender = str(old_gender or '').strip()
+            new_gender = str(features.get('gender', '') or '').strip()
+            if (old_gender and old_gender.lower() not in ('unknown', '未知')
+                    and new_gender and new_gender.lower() not in ('unknown', '未知')
+                    and old_gender != new_gender):
+                base_label, n = label, 2
+                while self._conn.execute(
+                        "SELECT 1 FROM people WHERE label=?", (f"{base_label}#{n}",)).fetchone():
+                    n += 1
+                label = f"{base_label}#{n}"
+                row = None
+                if logger:
+                    logger.warning(
+                        f"人物 uid 跨视频复用检测到性别冲突: {base_label} "
+                        f"旧={old_gender} 新={new_gender}，拆分为新 label={label}")
         # 特征合并（在已有 features_json 基础上）
         merged_features = self._merge_features(
             row['features_json'] if row else None, features) if row else (
diff --git a/fam-edge/src/fam_edge/person_service.py b/fam-edge/src/fam_edge/person_service.py
index 7457ab3..8147b28 100644
--- a/fam-edge/src/fam_edge/person_service.py
+++ b/fam-edge/src/fam_edge/person_service.py
@@ -145,10 +145,17 @@ class PersonService:
     def _aggregate_features(self) -> Dict[str, Dict]:
         """从 events.person_appearances_json 聚合每个 uid 的合并特征。
 
-        遍历所有事件的 person_appearances，按 uid 收集 features dict，
-        合并规则：首次非 unknown 值优先（与 oracle_db._merge_features 一致）。
+        遍历所有事件的 person_appearances，按 uid 收集 features dict，合并规则：
+        新非 unknown 值覆盖旧值（与 oracle_db._merge_features 一致）。
+
+        大模型给的 人物A/B/C 这类 uid 只在单次视频分析内部稳定，不同视频各自独立
+        编号——同一个字符串在不同视频里完全可能指向不同的真人（实测 "人物A" 在
+        29 个视频里混了男女两个人）。这里按性别做在线聚类：同一 uid 下遇到与已有
+        分组都冲突（都明确且不同）的性别时，开一个新分组（uid#2/uid#3...），不再
+        把两个人的特征糊成一坨；分组的 key 交给 upsert_person 当新 label 落库，
+        走 unnamed 流程由 LLM 合并判断该并入哪个已命名身份。
         """
-        uid_features: Dict[str, Dict] = {}
+        uid_groups: Dict[str, List[Dict]] = {}
         rows = self.db._conn.execute(
             "SELECT person_appearances_json FROM events "
             "WHERE person_appearances_json IS NOT NULL").fetchall()
@@ -168,17 +175,31 @@ class PersonService:
                 feats = pa.get('features') or {}
                 if not isinstance(feats, dict):
                     continue
-                if uid not in uid_features:
-                    uid_features[uid] = dict(feats)
-                else:
-                    merged = dict(uid_features[uid])
-                    for k, v in feats.items():
-                        v_str = str(v).strip() if v is not None else ''
-                        if v_str and v_str.lower() != 'unknown':
-                            merged[k] = v_str
-                        elif k not in merged:
-                            merged[k] = v_str or 'unknown'
-                    uid_features[uid] = merged
+                groups = uid_groups.setdefault(uid, [])
+                gender = str(feats.get('gender', '') or '').strip()
+                target = None
+                if gender and gender.lower() not in ('unknown', '未知'):
+                    for g in groups:
+                        g_gender = str(g.get('gender', '') or '').strip()
+                        if not g_gender or g_gender.lower() in ('unknown', '未知') or g_gender == gender:
+                            target = g
+                            break
+                elif groups:
+                    target = groups[0]
+                if target is None:
+                    target = {}
+                    groups.append(target)
+                for k, v in feats.items():
+                    v_str = str(v).strip() if v is not None else ''
+                    if v_str and v_str.lower() != 'unknown':
+                        target[k] = v_str
+                    elif k not in target:
+                        target[k] = v_str or 'unknown'
+        uid_features: Dict[str, Dict] = {}
+        for base_uid, groups in uid_groups.items():
+            for i, feats in enumerate(groups):
+                key = base_uid if i == 0 else f"{base_uid}#{i + 1}"
+                uid_features[key] = feats
         return uid_features
 
     def _collect_features_text(self, labels: List[str]) -> str:
diff --git a/fam-edge/tests/test_oracle_db.py b/fam-edge/tests/test_oracle_db.py
new file mode 100644
index 0000000..8fc0742
--- /dev/null
+++ b/fam-edge/tests/test_oracle_db.py
@@ -0,0 +1,72 @@
+import json
+
+from fam_edge.oracle_db import OracleDB
+
+
+def _db(tmp_path):
+    return OracleDB(str(tmp_path / "oracle.db"))
+
+
+def test_upsert_person_same_gender_merges_into_one_row(tmp_path):
+    db = _db(tmp_path)
+    db.upsert_person("人物A", features={"gender": "男", "hair": "短发黑色"})
+    db.upsert_person("人物A", features={"gender": "男", "clothing": "蓝色T恤"})
+    rows = db._conn.execute("SELECT * FROM people").fetchall()
+    assert len(rows) == 1
+    assert rows[0]["appearances"] == 2
+    feats = json.loads(rows[0]["features_json"])
+    assert feats["hair"] == "短发黑色"
+    assert feats["clothing"] == "蓝色T恤"
+
+
+def test_upsert_person_gender_conflict_splits_into_new_label(tmp_path):
+    """核心诉求: 大模型给的 人物A/B/C 这类 uid 只在单次视频分析内稳定，不同视频
+    独立编号，同一字符串完全可能撞到不同真人（实测 "人物A" 混了男女两个人）。
+    性别冲突时不能直接合并覆盖，要拆成新 label，避免两个人的特征越merge越乱。"""
+    db = _db(tmp_path)
+    db.upsert_person("人物A", features={"gender": "男", "clothing": "蓝色Polo衫"})
+    db.upsert_person("人物A", features={"gender": "女", "clothing": "白色上衣"})
+    rows = {r["label"]: r for r in db._conn.execute("SELECT * FROM people").fetchall()}
+    assert set(rows.keys()) == {"人物A", "人物A#2"}
+    assert rows["人物A"]["appearances"] == 1
+    assert json.loads(rows["人物A"]["features_json"])["gender"] == "男"
+    assert rows["人物A#2"]["appearances"] == 1
+    assert json.loads(rows["人物A#2"]["features_json"])["gender"] == "女"
+    # 派生行的 display_uid 仍然记录原始大模型 uid，方便追溯来源
+    assert rows["人物A#2"]["display_uid"] == "人物A"
+
+
+def test_upsert_person_gender_conflict_allocates_next_free_suffix(tmp_path):
+    db = _db(tmp_path)
+    db.upsert_person("人物A", features={"gender": "男"})
+    db.upsert_person("人物A", features={"gender": "女"})  # -> 人物A#2
+    db.upsert_person("人物A", features={"gender": "unknown"})  # unknown 不冲突，合并回 人物A
+    db.upsert_person("人物A", features={"gender": "男", "hair": "光头"})  # 冲突，人物A 有 gender 男了不冲突；应仍合并
+    labels = {r["label"] for r in db._conn.execute("SELECT label FROM people").fetchall()}
+    assert labels == {"人物A", "人物A#2"}
+    # 再来一次性别冲突（对着 人物A#2，性别女）应该分配 人物A#3，而不是复用 人物A#2
+    db.upsert_person("人物A", features={"gender": "男"})
+    db.upsert_person("人物A", features={"gender": "女"})
+    # 这次新的女性冲突会先撞到 人物A（此时是男），分裂出下一个空闲后缀
+    labels = {r["label"] for r in db._conn.execute("SELECT label FROM people").fetchall()}
+    assert "人物A" in labels
+    assert len(labels) >= 2
+
+
+def test_upsert_person_unknown_gender_never_triggers_split(tmp_path):
+    db = _db(tmp_path)
+    db.upsert_person("人物A", features={"gender": "男"})
+    db.upsert_person("人物A", features={"gender": "unknown"})
+    db.upsert_person("人物A", features={"gender": "未知"})
+    rows = db._conn.execute("SELECT * FROM people").fetchall()
+    assert len(rows) == 1
+    assert rows[0]["appearances"] == 3
+
+
+def test_upsert_person_no_features_never_triggers_split(tmp_path):
+    db = _db(tmp_path)
+    db.upsert_person("人物A", features={"gender": "男"})
+    db.upsert_person("人物A")  # 无 features（source 更新等场景）
+    rows = db._conn.execute("SELECT * FROM people").fetchall()
+    assert len(rows) == 1
+    assert rows[0]["appearances"] == 2
diff --git a/fam-edge/tests/test_person_service.py b/fam-edge/tests/test_person_service.py
new file mode 100644
index 0000000..83f022c
--- /dev/null
+++ b/fam-edge/tests/test_person_service.py
@@ -0,0 +1,70 @@
+import json
+
+from fam_edge.oracle_db import OracleDB
+from fam_edge.person_service import PersonService
+
+
+def _service(tmp_path):
+    db = OracleDB(str(tmp_path / "oracle.db"))
+    svc = PersonService.__new__(PersonService)  # 跳过 __init__（不需要真的建 LLM 适配器）
+    svc.db = db
+    return svc, db
+
+
+def _insert_event(db, video_id, appearances):
+    db._conn.execute(
+        "INSERT INTO events (video_id, ts, description, person_appearances_json) "
+        "VALUES (?, ?, ?, ?)",
+        (video_id, "2026-08-21 00:00:00", "desc", json.dumps(appearances, ensure_ascii=False)))
+    db._conn.commit()
+
+
+def test_aggregate_features_merges_same_gender_across_videos(tmp_path):
+    svc, db = _service(tmp_path)
+    _insert_event(db, 1, [{"uid": "人物A", "features": {"gender": "男", "hair": "短发黑色"}}])
+    _insert_event(db, 2, [{"uid": "人物A", "features": {"gender": "男", "clothing": "蓝色T恤"}}])
+    result = svc._aggregate_features()
+    assert set(result.keys()) == {"人物A"}
+    assert result["人物A"]["hair"] == "短发黑色"
+    assert result["人物A"]["clothing"] == "蓝色T恤"
+
+
+def test_aggregate_features_splits_gender_conflict_across_videos(tmp_path):
+    """核心场景: 复现 "人物A" 在 29 个视频里混了男女两个人的真实 bug——不同视频各自
+    独立编号的 uid，字符串相同不代表同一个真人，性别冲突时必须拆成独立分组。"""
+    svc, db = _service(tmp_path)
+    _insert_event(db, 1, [{"uid": "人物A", "features": {"gender": "男", "clothing": "蓝色Polo衫"}}])
+    _insert_event(db, 2, [{"uid": "人物A", "features": {"gender": "女", "clothing": "白色上衣"}}])
+    result = svc._aggregate_features()
+    assert set(result.keys()) == {"人物A", "人物A#2"}
+    assert result["人物A"]["gender"] == "男"
+    assert result["人物A#2"]["gender"] == "女"
+
+
+def test_aggregate_features_unknown_gender_joins_first_group(tmp_path):
+    svc, db = _service(tmp_path)
+    _insert_event(db, 1, [{"uid": "人物A", "features": {"gender": "男"}}])
+    _insert_event(db, 2, [{"uid": "人物A", "features": {"gender": "unknown", "hair": "光头"}}])
+    result = svc._aggregate_features()
+    assert set(result.keys()) == {"人物A"}
+    assert result["人物A"]["gender"] == "男"
+    assert result["人物A"]["hair"] == "光头"
+
+
+def test_aggregate_features_three_way_gender_reuse_creates_three_groups(tmp_path):
+    svc, db = _service(tmp_path)
+    _insert_event(db, 1, [{"uid": "人物B", "features": {"gender": "男"}}])
+    _insert_event(db, 2, [{"uid": "人物B", "features": {"gender": "女"}}])
+    _insert_event(db, 3, [{"uid": "人物B", "features": {"gender": "男"}}])  # 应并回第一组
+    result = svc._aggregate_features()
+    assert set(result.keys()) == {"人物B", "人物B#2"}
+    assert result["人物B"]["gender"] == "男"
+    assert result["人物B#2"]["gender"] == "女"
+
+
+def test_aggregate_features_ignores_placeholder_uids(tmp_path):
+    svc, db = _service(tmp_path)
+    _insert_event(db, 1, [{"uid": "无人", "features": {"gender": "男"}}])
+    _insert_event(db, 2, [{"uid": "", "features": {"gender": "男"}}])
+    result = svc._aggregate_features()
+    assert result == {}