diff --git a/cloud/tests/test_cloud_dispatch_pr.py b/cloud/tests/test_cloud_dispatch_pr.py
index 7427e34..e3cd22d 100644
--- a/cloud/tests/test_cloud_dispatch_pr.py
+++ b/cloud/tests/test_cloud_dispatch_pr.py
@@ -458,5 +458,26 @@ class ProjectTaskFieldTestCase(_ServerCase):
         self.assertEqual(body["executor"], "docker")
 
 
+class WorkerTablePrecisionTestCase(unittest.TestCase):
+    def test_workers_time_columns_are_double_precision(self):
+        # 回归：2026-10-26 主机实测，workers.last_seen 用 REAL 时，
+        # Postgres 的 float4 把 epoch 秒舍入到 128 秒一档，存活判断
+        # 误差可达 ±64 秒、甚至出现负 age。新表必须是 DOUBLE PRECISION。
+        import sqlite3 as _sqlite3
+        tmp = tempfile.TemporaryDirectory()
+        db_path = os.path.join(tmp.name, "w.db")
+        from cloud.workers import ensure_workers_table
+        ensure_workers_table(db_path)
+        conn = _sqlite3.connect(db_path)
+        try:
+            rows = conn.execute("PRAGMA table_info(workers)").fetchall()
+        finally:
+            conn.close()
+            tmp.cleanup()
+        by_name = {r[1]: r[2] for r in rows}
+        self.assertIn("DOUBLE", by_name["last_seen"])
+        self.assertIn("DOUBLE", by_name["created_at"])
+
+
 if __name__ == "__main__":
     unittest.main()
diff --git a/cloud/workers.py b/cloud/workers.py
index fb6a789..73d16f9 100644
--- a/cloud/workers.py
+++ b/cloud/workers.py
@@ -42,7 +42,19 @@ def alive_threshold_seconds(override=None):
 
 
 def ensure_workers_table(db_path):
-    """建 workers 表（不存在时建，已存在时不动）。"""
+    """建 workers 表（不存在时建，已存在时不动，老库补精度）。
+
+    注意：last_seen/created_at 必须是 DOUBLE PRECISION，不能用 REAL。
+    SQLite 的 REAL 是 8 字节双精度，没问题；但 Postgres 的 REAL 是
+    4 字节 float4，只能存约 7 位有效数字，epoch 秒（10 位）会被
+    舍入到 128 秒一档，导致存活判断误差可达 ±64 秒、甚至出现负 age
+    （2026-10-26 主机实测：last_seen 比 NOW() 还大 20 秒）。
+    新表直接用 DOUBLE PRECISION；老库（已是 REAL 列）用 ALTER
+    TYPE 升到 DOUBLE PRECISION（float4 -> float8 无损，已存的舍入值
+    无法恢复，但之后的心跳不再舍入）。SQLite 下 ALTER TYPE 语法
+    不支持，靠 try/except 跳过——SQLite 的 REAL 本来就是双精度，
+    不需要升。
+    """
     conn = _db_connect(str(db_path))
     try:
         conn.execute(
@@ -50,10 +62,38 @@ def ensure_workers_table(db_path):
             "worker_id TEXT PRIMARY KEY, "
             "executor TEXT, "
             "capacity INTEGER, "
-            "last_seen REAL, "
-            "created_at REAL)"
+            "last_seen DOUBLE PRECISION, "
+            "created_at DOUBLE PRECISION)"
         )
         conn.commit()
+        # 老库升精度只在确实是 Postgres REAL 列时做一次：每次心跳都
+        # ALTER 会反复拿 ACCESS EXCLUSIVE 锁，没必要。先查
+        # information_schema（SQLite 下这张表不存在，查询失败即跳过，
+        # SQLite 的 REAL 本来就是双精度，不需要升）。
+        for col in ("last_seen", "created_at"):
+            try:
+                row = conn.execute(
+                    "SELECT data_type FROM information_schema.columns "
+                    "WHERE table_name = 'workers' AND column_name = ?",
+                    (col,)).fetchone()
+            except Exception:
+                try:
+                    conn.rollback()
+                except Exception:
+                    pass
+                continue
+            if not row or row[0] != "real":
+                continue
+            try:
+                conn.execute(
+                    "ALTER TABLE workers ALTER COLUMN %s "
+                    "TYPE DOUBLE PRECISION" % col)
+                conn.commit()
+            except Exception:
+                try:
+                    conn.rollback()
+                except Exception:
+                    pass
     finally:
         conn.close()