Keep register pool monitor alive across transient errors

The available/quota-mode register loop runs a background monitor thread
that re-checks the account pool every check_interval and resumes
registering when it drops below target. Two issues made that thread die
silently, leaving enabled=True (UI still shows "运行中") while nothing
re-evaluated the pool — only a manual stop→start recovered it:

- _run had no per-iteration try/except, so any transient error in the
loop body killed the whole thread.
- _bump persisted register.json to the /data FUSE bucket on every pool
check (~2 writes / interval). A single bucket write hiccup raised and
unwound the unguarded loop — the exact per-write-to-FUSE anti-pattern
CLAUDE.md already warns against.

Fixes:
- Wrap the loop body in try/except (log + sleep + continue); add a
finally that always sets enabled=False so the UI can't get stuck on a
dead thread.
- Drop _save() from _bump. Stats are ephemeral display state read from
in-memory _config over SSE and rebuilt by start(); persist only on
start/stop/update/reset.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

services/register_service.py

@@ -183,38 +183,58 @@ class RegisterService:
                 stats["avg_seconds"] = round(elapsed / success, 1) if success else 0
                 stats["success_rate"] = round(success * 100 / max(1, success + fail), 1)
             self._config["stats"]["updated_at"] = _now()
-            self._save()
+            # 不在此落盘：stats 只是临时展示字段，SSE 读内存 _config 即可，start() 每次重建。
+            # _bump 每次号池检查都调用，旧实现会把 register.json 写到 /data FUSE 桶；
+            # 桶写一次抽风抛异常，就会杀死没有 try 的 _run 监控线程，enabled 卡在 True、
+            # 号池掉到目标下也不再触发注册（CLAUDE.md 已警告对象桶每次全量写的反模式）。
+            # 真正落盘只发生在 start/stop/update/reset。
 
     def _run(self) -> None:
         threads = int(self.get()["threads"])
         submitted, done, success, fail = 0, 0, 0, 0
-        with ThreadPoolExecutor(max_workers=threads) as executor:
-            futures = set()
-            while True:
-                cfg = self.get()
-                while self.get()["enabled"] and not self._target_reached(cfg, submitted) and len(futures) < threads:
-                    submitted += 1
-                    futures.add(executor.submit(openai_register.worker, submitted))
-                self._bump(running=len(futures), done=done, success=success, fail=fail)
-                if not futures and (not self.get()["enabled"] or str(cfg.get("mode") or "total") == "total"):
-                    break
-                if not futures:
-                    time.sleep(max(1, int(cfg.get("check_interval") or 5)))
-                    continue
-                finished, futures = wait(futures, return_when=FIRST_COMPLETED)
-                for future in finished:
-                    done += 1
+        try:
+            with ThreadPoolExecutor(max_workers=threads) as executor:
+                futures: set = set()
+                while True:
                     try:
-                        result = future.result()
-                        success += 1 if result.get("ok") else 0
-                        fail += 0 if result.get("ok") else 1
-                    except Exception:
-                        fail += 1
-        self._bump(running=0, done=done, success=success, fail=fail, finished_at=_now())
-        with self._lock:
-            self._config["enabled"] = False
-            self._save()
-        self._append_log(f"注册任务结束，成功{success}，失败{fail}", "yellow")
+                        cfg = self.get()
+                        while self.get()["enabled"] and not self._target_reached(cfg, submitted) and len(futures) < threads:
+                            submitted += 1
+                            futures.add(executor.submit(openai_register.worker, submitted))
+                        self._bump(running=len(futures), done=done, success=success, fail=fail)
+                        if not futures and (not self.get()["enabled"] or str(cfg.get("mode") or "total") == "total"):
+                            break
+                        if not futures:
+                            time.sleep(max(1, int(cfg.get("check_interval") or 5)))
+                            continue
+                        finished, futures = wait(futures, return_when=FIRST_COMPLETED)
+                        for future in finished:
+                            done += 1
+                            try:
+                                result = future.result()
+                                success += 1 if result.get("ok") else 0
+                                fail += 0 if result.get("ok") else 1
+                            except Exception:
+                                fail += 1
+                    except Exception as loop_error:
+                        # 单次循环出错（如 /data FUSE 桶写盘抽风、号池读取瞬时异常）绝不能杀死
+                        # 监控线程：否则 enabled 卡在 True、徽章仍显示“运行中”，但号池掉到目标下
+                        # 也永不再触发注册，只能手动 stop→start 才恢复。忽略本次，睡一个检查间隔后继续。
+                        self._append_log(f"号池监控循环异常，已忽略本次并继续: {loop_error}", "red")
+                        time.sleep(max(1, int(self.get().get("check_interval") or 5)))
+        finally:
+            # 无论正常结束还是异常退出，都把 enabled 落为 False，避免 UI 卡在“运行中”假象。
+            try:
+                self._bump(running=0, done=done, success=success, fail=fail, finished_at=_now())
+            except Exception:
+                pass
+            with self._lock:
+                self._config["enabled"] = False
+                try:
+                    self._save()
+                except Exception:
+                    pass
+            self._append_log(f"注册任务结束，成功{success}，失败{fail}", "yellow")
 
 
 register_service = RegisterService(REGISTER_FILE)