diff --git a/tests/test_health.py b/tests/test_health.py
index 5e1df05..9ae03d4 100644
--- a/tests/test_health.py
+++ b/tests/test_health.py
@@ -4,6 +4,7 @@
 
 import threading
 import time
+from collections.abc import Callable
 
 import pytest
 
@@ -158,32 +159,48 @@ def test_remote_check_disabled() -> None:
 
 
 def test_run_loop_stops_cleanly() -> None:
-    # Use a healthy manager: with backoff enabled an unhealthy one would grow
-    # the interval and could starve the ">= 3 checks" assertion below.
+    """循环反复检查，而设置停止事件后线程真的结束了。
+
+    不按墙钟断言："睡 0.3 秒后至少 3 次检查"考的是**这台机器够快**，而 CI 上带
+    coverage 插桩的 macOS runner 把 ``run_loop`` 的一轮从标称 50ms 拉到 120ms，
+    于是等到 2 次就断言失败（main 上真实发生过）。这里改成：次数用带截止时间的
+    等待，"停下来"则用 join 证明——那才是测试名字里的 stops cleanly。
+    """
+    # 用健康的 manager：不健康会触发退避，间隔指数增长，把"若干次检查"饿死。
     hc = HealthChecker(_TM(alive=True, ports="list"), _hc())
     seen: list[HealthStatus] = []
+    before_threads = set(threading.enumerate())
     stop = hc.run_loop(interval=0.05, callback=lambda st: seen.append(st))
     assert isinstance(stop, threading.Event)
-    time.sleep(0.3)
+    monitor = next((t for t in threading.enumerate() if t not in before_threads), None)
+    assert monitor is not None, "run_loop 没有起后台线程"
+
+    _wait_for_values(seen, 3)
+
+    checks_at_stop = len(seen)
     stop.set()
-    time.sleep(0.15)
-    assert len(seen) >= 3, len(seen)
+    monitor.join(timeout=5)
+    assert not monitor.is_alive(), "设置停止事件后循环没有退出"
+    # 已经进入 check() 的那一轮仍会回调一次（实现只保证在 wait 之前查一次事件），
+    # 再多就说明停止没生效。线程已 join，计数此后不会再变，所以这是确定性的。
+    late = len(seen) - checks_at_stop
+    assert late <= 1, f"停止后又回调了 {late} 次"
     assert all(isinstance(st, HealthStatus) for st in seen)
-    n = len(seen)
-    time.sleep(0.15)
-    assert len(seen) == n, (len(seen), n)
 
 
 def test_run_loop_callback_error_swallowed() -> None:
+    """回调抛出的异常被记下来，而不是把监控线程弄死。"""
     hc = HealthChecker(_TM(alive=True, ports="dict"), _hc())
 
     def bad_cb(_st: HealthStatus) -> None:
         raise ValueError("cb boom")
 
     stop = hc.run_loop(interval=0.02, callback=bad_cb)
-    time.sleep(0.15)
+    _wait_for(
+        lambda: isinstance(hc.last_callback_error, ValueError),
+        lambda: f"回调异常没被记录：{hc.last_callback_error!r}",
+    )
     stop.set()
-    assert isinstance(hc.last_callback_error, ValueError)
 
 
 def test_health_status_str() -> None:
@@ -281,12 +298,29 @@ def test_backoff_interval_formula() -> None:
     assert HealthChecker._backoff_interval(60.0, 10, 300.0) == 300.0
 
 
-def _wait_for_values(values: list, n: int, timeout: float = 3.0) -> None:
-    """Busy-wait until ``values`` has at least ``n`` entries."""
+def _wait_for(
+    predicate: Callable[[], bool],
+    describe: Callable[[], str],
+    timeout: float = 5.0,
+) -> None:
+    """Wait for *predicate* on a deadline, instead of napping a fixed time.
+
+    固定秒数的 ``sleep`` 断言的是"这台机器够快"，而不是被测行为。``describe``
+    是个函数：失败信息要在**等过之后**才取，否则它记的是等待前那一刻的状态。
+    """
     deadline = time.time() + timeout
-    while len(values) < n and time.time() < deadline:
+    while not predicate() and time.time() < deadline:
         time.sleep(0.01)
-    assert len(values) >= n, f"captured only {len(values)} values, need {n}"
+    assert predicate(), describe()
+
+
+def _wait_for_values(values: list, n: int, timeout: float = 5.0) -> None:
+    """Busy-wait until ``values`` has at least ``n`` entries."""
+    _wait_for(
+        lambda: len(values) >= n,
+        lambda: f"captured only {len(values)} values, need {n}",
+        timeout,
+    )
 
 
 def _fake_wait_recorder(