fix: PG init backfill rollback scope + audit-loop state upsert registration

- db: translate INSERT OR IGNORE ... SELECT to ON CONFLICT DO NOTHING
  (VALUES-only regex left SELECT form untranslated -> PG syntax error);
  register task_audit_loop_state (task_id) in _OR_REPLACE_PK so PG
  INSERT OR REPLACE becomes a real upsert instead of DO NOTHING.
- server: _backfill_default_quotas uses SAVEPOINT containment and no
  longer commits the caller's transaction, so a backfill failure cannot
  roll back init_db's migrate ALTERs.
- worker: _coord_child_rows rolls back before column-fallback retries so
  a missing-column failure on PG does not cascade into
  InFailedSqlTransaction.
- tests: test_pg_init_fix.py (translation, SQLite backfill/upsert,
  real-PG fresh-init + merge + upsert + backfill containment, skipped
  without CLOUD_TEST_PG_DSN/driver).

cloud/server.py

@@ -1642,8 +1642,28 @@ def _backfill_default_quotas(conn):
 
     只补“users 里有、user_quotas 里无”的用户，已有行（含
     管理员显式设为 NULL 的不限额行）一律不动。返回补写行数。
+
+    失败只回滚到保存点，绝不回滚调用方的整个事务：init_db
+    里 migrate 的 ALTER 与它同事务，旧版 except 里整事务
+    rollback 会把 ALTER 一并丢掉（2026-10-31 新鲜 PG 库实测：
+    init 后 tasks 缺 paused_by_user / interrupted_reason，合流
+    随后在 PG“语句失败即 abort 整个事务”的语义下连环报
+    InFailedSqlTransaction）。成功时也不再由本函数提交，提交
+    由调用方负责（init_db 结尾统一提交；旧版成功时提交等于
+    提前提交 init 的全部 DDL，改后 init 恢复原子性）。调用方
+    只有 init_db 两处（均以提交收尾），改动不影响其它路径。
+    保存点在 SQLite 与 PG 下都支持；保存点本身建不起来时
+    （如连接已处 abort 态）退化为整事务 rollback 再返回 0。
     """
     _ensure_quotas_table(conn)
+    try:
+        conn.execute("SAVEPOINT backfill_default_quotas")
+    except Exception:
+        try:
+            conn.rollback()
+        except Exception:
+            pass
+        return 0
     try:
         cur = conn.execute(
             "INSERT OR IGNORE INTO user_quotas(user_id,"
@@ -1652,13 +1672,17 @@ def _backfill_default_quotas(conn):
             " LEFT JOIN user_quotas q ON q.user_id = u.id"
             " WHERE q.user_id IS NULL",
             (_default_quota_usd_micros(), time.time()))
-        conn.commit()
+        conn.execute("RELEASE SAVEPOINT backfill_default_quotas")
         return cur.rowcount or 0
     except Exception:
         try:
-            conn.rollback()
+            conn.execute("ROLLBACK TO SAVEPOINT backfill_default_quotas")
+            conn.execute("RELEASE SAVEPOINT backfill_default_quotas")
         except Exception:
-            pass
+            try:
+                conn.rollback()
+            except Exception:
+                pass
         return 0