diff --git a/.gitignore b/.gitignore
index 4b5b46c..d19070e 100644
--- a/.gitignore
+++ b/.gitignore
@@ -22,6 +22,7 @@ media/
 
 # Local scratch DBs used for Alembic autogenerate (see migrations/env.py)
 *.db
+*.db.retired
 *.sqlite
 *.sqlite3
 .playwright-mcp/
diff --git a/yunke-server/scripts/backfill_group_members.py b/yunke-server/scripts/backfill_group_members.py
index 6852f5e..03f5676 100644
--- a/yunke-server/scripts/backfill_group_members.py
+++ b/yunke-server/scripts/backfill_group_members.py
@@ -26,13 +26,22 @@ from app.storage.repos import _upsert_one
 logging.basicConfig(level=logging.INFO, format="%(asctime)s %(levelname)s %(message)s")
 log = logging.getLogger("backfill_group_members")
 
-# max(timestamp) 裸列：SQLite 保证 sender_name 取自 timestamp 最大的那行
+# row_number() 窗口取 timestamp 最新一行的 sender_name：MySQL 8 的
+# only_full_group_by 拒绝 SQLite 式「max()+裸列」(1055)，窗口写法两库通用
+# 且语义严格（最新一条的昵称），不依赖方言碰巧行为
 SQL = sa_text("""
-SELECT room_id, talker_id, sender_name, max(timestamp) AS last_ts, count(*) AS msg_count
-FROM message
-WHERE channel = 'wechat' AND room_id IS NOT NULL AND room_id != ''
-  AND talker_id IS NOT NULL AND talker_id != ''
-GROUP BY room_id, talker_id
+SELECT room_id, talker_id, sender_name, last_ts, msg_count
+FROM (
+  SELECT room_id, talker_id, sender_name, timestamp AS last_ts,
+         count(*) OVER (PARTITION BY room_id, talker_id) AS msg_count,
+         row_number() OVER (
+           PARTITION BY room_id, talker_id ORDER BY timestamp DESC
+         ) AS rn
+  FROM message
+  WHERE channel = 'wechat' AND room_id IS NOT NULL AND room_id != ''
+    AND talker_id IS NOT NULL AND talker_id != ''
+) t
+WHERE rn = 1
 """)