diff --git a/backend/app/cluster/fence.py b/backend/app/cluster/fence.py
index ad18d85f..1d94b3d6 100644
--- a/backend/app/cluster/fence.py
+++ b/backend/app/cluster/fence.py
@@ -6,9 +6,10 @@
 
 - 续租约在单独的任务、单独的连接上做（cluster/service.py），不被同一拍别的活拖住；最近一次续上的时刻记在这里（renewed）。
 - 本进程的执行在动手写之前问一句 check()：离上次续上已经超过一个租约时长（别的进程按两个租约时长才接手，中间留着一段），或者这段
-  执行是在上一次隔离之前开始的（世代不对），就抛 Fenced——写事件、写断点、写终态、写待办和动作、发外部写入之前都问。
+  执行是在上一次隔离之前开始的（世代不对），就抛 Fenced——写事件、写断点、写终态、写待办和动作、发外部写入之前都问。只对绑定了
+  世代的执行（运行的一段执行、后台任务）抛；别的调用者问到时只隔离、不抛。
 - 一旦发现续晚了（trip）：世代加一，手上的执行全部停掉（停的时候不写终态，那些行交给接手的进程按「属主不在了」处理），换一个新的
-  身份重新登记，之后发起的照常。
+  身份重新登记；新身份登记上之前新开始的执行先等着（ready），登记上之后照常。
 
 **用哪只钟（复核 R-b7）。** 别的进程判「租约过期」用的是库的时钟（cluster/workers：续租写 lease_until = 库的 now() + 租约），
 跟各台机器的墙钟差多少无关。这边判「续晚了」用计入休眠的钟：time.monotonic() 在 macOS（mach_absolute_time）、Linux
@@ -21,6 +22,7 @@ Fenced 是 BaseException：节点、工具里常见的 except Exception 拦不
 """
 from __future__ import annotations
 
+import asyncio
 import contextvars
 import logging
 import sys
@@ -96,17 +98,38 @@ def bind() -> None:
 
 
 def stale() -> bool:
-    """离上次续上租约是不是已经超过一个租约时长：两只钟（计入休眠的单调钟、墙钟）里走得多的那只说了算。"""
+    """离上次续上租约是不是已经超过一个租约时长：两只钟（计入休眠的单调钟、墙钟）里走得多的那只说了算。隔离之后、新身份还没
+    登记上的那一段也算（pending）。"""
     if not _active:
         return False
     since = _since()
     if since is None:
-        return False
+        return True
     from app.cluster.workers import lease_s
 
     return max(since) > lease_s()
 
 
+def pending() -> bool:
+    """隔离过、换的新身份还没登记上。"""
+    return _active and _renewed is None
+
+
+async def ready(timeout: float | None = None) -> None:
+    """新开始的一段执行（运行、后台任务）在 bind() 之前等一等：隔离之后、新身份还没登记上时，它认领的行在别的进程眼里属主
+    不在，会被当成没人管接手（复核 R-b9）。最多等一个租约时长（timeout），等不到抛 Fenced——不开始，那几行由别的进程照「属主
+    不在了」处理（cluster/service 那边同时会放弃这个一直没登记上的身份）。"""
+    if not pending():
+        return
+    from app.cluster.workers import lease_s
+
+    deadline = _clock() + (lease_s() if timeout is None else timeout)
+    while pending():
+        if _clock() >= deadline:
+            raise Fenced()
+        await asyncio.sleep(0.05)
+
+
 def jumped() -> bool:
     """两只钟差出一个租约以上（墙钟往前或往后跳了，或者单调钟没计入休眠）：续租时问，是就隔离。"""
     if not _active:
@@ -120,22 +143,22 @@ def jumped() -> bool:
 
 
 def fenced_here() -> bool:
-    """当前这段执行还能不能写：世代不对（隔离之前开始的）或者租约续晚了。"""
+    """当前这段执行还能不能写：世代不对（隔离之前开始的）或者租约续晚了。只说绑定了世代的执行（bind 过的）。"""
     if not _active:
         return False
     bound = _bound.get()
-    return (bound is not None and bound != _generation) or stale()
+    return bound is not None and (bound != _generation or stale())
 
 
 def check() -> None:
-    """写之前问一句：不许写就抛 Fenced（续晚了的话顺手隔离）。"""
+    """写之前问一句。续晚了就隔离（谁先问到都一样）；抛 Fenced 只对绑定了世代的执行（运行的一段执行、后台任务：bind 过的）——
+    别的调用者（HTTP 处理、调度那一拍、进程间请求）不是认领了行在跑的执行，抛给它们只会把调度循环之类的打停（复核 R-b9）。"""
     if not _active:
         return
-    if stale():
+    if _renewed is not None and stale():
         trip("执行要写的时候发现租约已经续晚了")
-        raise Fenced()
     bound = _bound.get()
-    if bound is not None and bound != _generation:
+    if bound is not None and (bound != _generation or stale()):
         raise Fenced()
 
 
@@ -162,4 +185,4 @@ def trip(why: str) -> bool:
 
 
 __all__ = ["Fenced", "SLEEP_AWARE", "activate", "bind", "check", "deactivate", "fenced_here", "generation", "jumped", "now",
-           "on_trip", "renewed", "stale", "trip"]
+           "on_trip", "pending", "ready", "renewed", "stale", "trip"]
diff --git a/backend/app/cluster/requests.py b/backend/app/cluster/requests.py
index 34fb7e8f..2711d4d6 100644
--- a/backend/app/cluster/requests.py
+++ b/backend/app/cluster/requests.py
@@ -94,9 +94,11 @@ _running: set[asyncio.Future[Any]] = set()
 
 async def _handle(fn: Callable[..., Awaitable[Any]], kind: str, target: str, target_id: str, actor: str | None,
                   payload: dict[str, Any]) -> None:
+    from app.cluster.fence import Fenced
+
     try:
         await fn(target_id, actor, payload)
-    except Exception:  # noqa: BLE001
+    except (Exception, Fenced):  # noqa: BLE001 - Fenced 是 BaseException，不接住就成了没人取的任务异常
         logger.exception("办进程间请求 %s/%s %s 出错", kind, target, target_id)
 
 
diff --git a/backend/app/cluster/service.py b/backend/app/cluster/service.py
index 1fbd7ace..2d97a18b 100644
--- a/backend/app/cluster/service.py
+++ b/backend/app/cluster/service.py
@@ -65,6 +65,11 @@ class _Service:
         #: SQLite：守卫已经确认只有本进程——登记只写一次、不续租、本进程直接当班、接手只在启动时做（不为协调多写库）
         self.single = False
         self._lease_conn: Any = None
+        self.lease_wake: asyncio.Event | None = None
+        #: 隔离之后、新身份还没登记上：从什么时候起（cluster/fence.now 的读数）
+        self.pending_since: float | None = None
+        #: 隔离掉（或者一直没登记上、放弃了）的旧身份：新身份登记上之后收拾它们留下的登记、名额、请求、当班租约
+        self.retired: list[str] = []
 
     async def start(self) -> None:
         """登记、开始收发通知、起心跳。SQLite 时先取单进程的守卫（拿不到抛 guard.SingleProcessOnly）。"""
@@ -94,6 +99,7 @@ class _Service:
             await self._lead()
             fence.activate()
             fence.on_trip(self._on_fenced)
+            self.lease_wake = asyncio.Event()
             self.lease_task = asyncio.create_task(self._lease_loop(), name="agentlab-cluster-lease")
             self.task = asyncio.create_task(self._loop(), name="agentlab-cluster-heartbeat")
         logger.info("工作进程 %s 已登记（%s，租约 %g 秒、心跳 %g 秒；现在另有 %d 个进程）", identity.describe(identity.owner()),
@@ -159,27 +165,62 @@ class _Service:
         return True
 
     async def _lease_loop(self) -> None:
+        """续租约；隔离之后改成重新登记新身份。哪一下出了错（库一抖、版本守卫拦下）都不让这个任务退出——它一退，之后就再也没有
+        隔离保护了（复核 R-b9）：照实记日志，退避着重试（最长一个租约时长）。"""
+        from app.cluster import fence
+
+        delay = workers.heartbeat_s()
         while True:
-            await self.renew()
-            await asyncio.sleep(workers.heartbeat_s())
+            try:
+                if self.pending_since is not None:
+                    await self.register_again()
+                else:
+                    await self.renew()
+                delay = workers.heartbeat_s()
+            except asyncio.CancelledError:
+                raise
+            except (Exception, fence.Fenced):  # noqa: BLE001
+                delay = min(max(delay * 2, workers.heartbeat_s()), workers.lease_s())
+                logger.exception("%s出错，%g 秒后再试", "隔离之后重新登记" if self.pending_since is not None else "续租约", delay)
+            if self.lease_wake is None:
+                await asyncio.sleep(delay)
+                continue
+            with contextlib.suppress(asyncio.TimeoutError):
+                await asyncio.wait_for(self.lease_wake.wait(), timeout=delay)
+            self.lease_wake.clear()
 
     def _on_fenced(self) -> None:
-        """隔离了：手上的执行由各处自己停（runner、job_runtime 登记的处理），这里换身份重新登记。"""
-        asyncio.ensure_future(self._reidentify())
-
-    async def _reidentify(self) -> None:
+        """隔离了：手上的执行由各处自己停（runner、job_runtime 登记的处理）；这里马上换身份（之后新建的行记新身份），叫续租的任务
+        去登记它。"""
         from app.cluster import fence
 
         old = identity.owner()
+        self.retired.append(old)
         new = identity.regenerate()
+        self.pending_since = fence.now()
         leader.scheduler.drop()
-        try:
-            await workers.register()
-            fence.renewed()
-            logger.warning("隔离之后换了身份重新登记：%s → %s（之前认领的行由当班进程按「属主不在了」接手）",
-                           identity.describe(old), identity.describe(new))
-        except Exception:  # noqa: BLE001
-            logger.exception("隔离之后重新登记失败，下一拍再试")
+        logger.warning("隔离：换成新身份 %s（旧身份 %s），登记上之前新开始的执行先等着", new, old)
+        if self.lease_wake is not None:
+            self.lease_wake.set()
+
+    async def register_again(self) -> None:
+        """隔离之后登记新身份。一个租约时长都没登记上（库连不上、版本守卫拦着）就放弃这个身份、再换一个：这期间等不到的执行已经
+        放弃开始（fence.ready），它们的行记着这个身份——它永远不登记，别的进程就照「属主不在了」处理那几行。"""
+        from app.cluster import fence
+
+        assert self.pending_since is not None
+        if fence.now() - self.pending_since > workers.lease_s():
+            gave_up = identity.owner()
+            self.retired.append(gave_up)
+            identity.regenerate()
+            self.pending_since = fence.now()
+            logger.error("新身份 %s 一个租约时长都没登记上，放弃它、换成 %s 再试", gave_up, identity.owner())
+        await workers.register()
+        retired, self.retired = self.retired, []
+        self.pending_since = None
+        fence.renewed()
+        logger.warning("隔离之后换了身份重新登记：%s → %s（之前认领的行由当班进程按「属主不在了」接手）",
+                       "、".join(retired), identity.owner())
 
     # ---------------- 当班、心跳的其余活 ----------------
 
@@ -203,8 +244,9 @@ class _Service:
         return now_leading
 
     async def beat(self) -> None:
-        """心跳的一拍（续租约另在 _lease_loop；测试里直接调）。SQLite 只做本进程自己攒着的活。"""
-        if self.single:
+        """心跳的一拍（续租约另在 _lease_loop；测试里直接调）。SQLite 只做本进程自己攒着的活；隔离之后新身份还没登记上时也只做
+        这些（不争当班、不接手）。"""
+        if self.single or self.pending_since is not None:
             await self._own_work()
             return
         live = await workers.liveness()
@@ -254,6 +296,8 @@ class _Service:
         return out
 
     async def _loop(self) -> None:
+        from app.cluster import fence
+
         assert self.wake is not None
         while True:
             with contextlib.suppress(asyncio.TimeoutError):
@@ -263,7 +307,7 @@ class _Service:
                 await self.beat()
             except asyncio.CancelledError:
                 raise
-            except Exception:  # noqa: BLE001 - 一拍出错（库一时不通）不该让心跳停下
+            except (Exception, fence.Fenced):  # noqa: BLE001 - 一拍出错（库一时不通）不该让心跳停下
                 logger.exception("心跳这一拍出错，下一拍照常进行")
 
     async def _light_loop(self) -> None:
@@ -284,6 +328,7 @@ class _Service:
                 with contextlib.suppress(asyncio.CancelledError, Exception):
                     await task
         self.task = self.lease_task = None
+        self.pending_since, self.retired = None, []
         fence.deactivate()
         with contextlib.suppress(Exception):
             await leader.scheduler.release()
diff --git a/backend/app/core/job_runtime.py b/backend/app/core/job_runtime.py
index b050ee83..15775ee7 100644
--- a/backend/app/core/job_runtime.py
+++ b/backend/app/core/job_runtime.py
@@ -543,6 +543,11 @@ class JobRun:
     async def run(self) -> None:
         from app.cluster import fence
 
+        try:
+            await fence.ready()   # 隔离之后新身份还没登记上时先等一等（cluster/fence.ready）
+        except fence.Fenced:
+            logger.error("%s %s：本进程隔离之后一直没重新登记上，不开始（由别的进程按「属主不在了」处理）", self.kind, self.job_id)
+            return
         fence.bind()
         try:
             try:
diff --git a/backend/app/engine/runner.py b/backend/app/engine/runner.py
index 141b9b41..5d4c4af7 100644
--- a/backend/app/engine/runner.py
+++ b/backend/app/engine/runner.py
@@ -794,7 +794,12 @@ class RunManager:
         # 等待从 run.interrupted 算起——以前执行一直计到 finally，收尾那一截执行和等待两边都算，
         # active + wait 比墙钟还长（机器一忙就差出几十毫秒）
         paused: float | None = None
-        # 这段执行属于现在这一世代（cluster/fence.py）：本进程隔离之后，它不再写
+        # 这段执行属于现在这一世代（cluster/fence.py）：本进程隔离之后，它不再写。隔离之后新身份还没登记上时先等一等
+        try:
+            await fence.ready()
+        except Fenced:
+            logger.error("运行 %s：本进程隔离之后一直没重新登记上，不开始执行（由别的进程按「属主不在了」处理）", run_id)
+            return
         fence.bind()
         mine = identity.owner()
         #: 被隔离了（续不上租约）：停手，不写终态（这次运行由接手的进程按「属主不在了」处理）
diff --git a/backend/app/scheduler/loop.py b/backend/app/scheduler/loop.py
index 7a7fba52..6c2a9f14 100644
--- a/backend/app/scheduler/loop.py
+++ b/backend/app/scheduler/loop.py
@@ -201,6 +201,7 @@ class _Runner:
 
     async def _loop(self, interval: float, stopping: asyncio.Event) -> None:
         from app.cluster import leading
+        from app.cluster.fence import Fenced
         from app.scheduler import webhook
 
         self._poked = asyncio.Event()
@@ -218,7 +219,7 @@ class _Runner:
                     self.notify = asyncio.create_task(_notify(webhook), name="agentlab-scheduler-notify")
             except asyncio.CancelledError:
                 raise
-            except Exception:  # noqa: BLE001 - 这一拍出错不该让循环停下
+            except (Exception, Fenced):  # noqa: BLE001 - 这一拍出错不该让循环停下（Fenced 是 BaseException：复核 R-b9）
                 logger.exception("定时任务调度这一拍出错，下一拍照常进行")
             await self._wait(stopping, interval)
 
diff --git a/backend/tests/test_cluster_fence.py b/backend/tests/test_cluster_fence.py
index 18f99fd1..5d3908bb 100644
--- a/backend/tests/test_cluster_fence.py
+++ b/backend/tests/test_cluster_fence.py
@@ -4,7 +4,7 @@
 from __future__ import annotations
 
 import asyncio
-import time
+import contextlib
 
 import pytest
 from sqlalchemy import func, select
@@ -168,3 +168,108 @@ async def test_a_repeated_seq_is_refused_by_the_database():
                          RunEvent(id=new_id(), run_id=run.id, seq=1, type="log", ts=0, data={})])
         with pytest.raises(IntegrityError):
             await session.commit()
+
+
+# ---------------- 复核 R-b9：谁会被打停、谁不会 ----------------
+
+
+async def test_only_executions_bound_to_a_generation_are_stopped_by_check():
+    """续晚之后第一个问 check() 的不一定是运行本身：没绑定世代的调用者（HTTP 处理、调度那一拍）只隔离、不抛；绑定了的抛。
+    隔离之后新身份还没登记上，新开始的执行在 ready() 上等，等不到抛 Fenced；登记上了就放行。"""
+    gate = asyncio.Event()
+    seen: list[str] = []
+
+    async def bound_writer() -> None:
+        fence.bind()
+        await gate.wait()
+        try:
+            fence.check()
+            seen.append("wrote")
+        except fence.Fenced:
+            seen.append("fenced")
+
+    task = asyncio.create_task(bound_writer())
+    await asyncio.sleep(0)
+    before = fence.generation()
+    fence.renewed(ago=lease_s() + 1)
+    fence.check()                                      # 没绑定世代：只隔离
+    assert fence.generation() == before + 1 and fence.pending()
+    gate.set()
+    await task
+    assert seen == ["fenced"], "绑定在隔离之前的那一段执行：不许再写"
+
+    with pytest.raises(fence.Fenced):
+        await fence.ready(timeout=0.1)                 # 新身份还没登记上：新开始的执行等不到就不开始
+    waiting = asyncio.create_task(fence.ready(timeout=5))
+    await asyncio.sleep(0.1)
+    assert not waiting.done()
+    fence.renewed()                                    # 登记上了
+    await asyncio.wait_for(waiting, timeout=2)
+    fence.bind()
+    fence.check()                                      # 新世代照常
+
+
+async def test_the_scheduler_loop_survives_a_fenced_tick(monkeypatch):
+    from app.scheduler import loop as sloop
+
+    calls = {"n": 0}
+
+    async def fire_due(now=None):
+        calls["n"] += 1
+        if calls["n"] == 1:
+            raise fence.Fenced()
+        return []
+
+    monkeypatch.setattr(sloop, "fire_due", fire_due)
+    await sloop.runner.start(interval=0.02)
+    try:
+        for _ in range(200):
+            if calls["n"] >= 3:
+                break
+            await asyncio.sleep(0.02)
+        assert calls["n"] >= 3 and sloop.runner.running(), "被 Fenced 打到的那一拍之后照常往下"
+    finally:
+        await sloop.runner.stop()
+
+
+async def test_the_lease_task_keeps_going_when_registering_again_fails(monkeypatch):
+    """隔离之后重新登记：库一抖、版本守卫拦下都不让续租的任务退出，退避着重试；登记上了照常续租。"""
+    from app.cluster import identity, service as service_mod, workers
+    from app.cluster.version import VersionMismatch
+
+    monkeypatch.setattr(identity, "OWNER", identity.OWNER)    # 测试完换回原来的身份
+    monkeypatch.setattr(workers, "heartbeat_s", lambda: 0.01)
+    monkeypatch.setattr(workers, "lease_s", lambda: 5.0)
+    svc = service_mod._Service()
+    tries = {"register": 0, "renew": 0}
+
+    async def register(*, exclusive: bool = False) -> None:  # noqa: ARG001
+        tries["register"] += 1
+        if tries["register"] == 1:
+            raise VersionMismatch("库里还有版本不同的进程活着")
+        if tries["register"] == 2:
+            raise OSError("库一时连不上")
+
+    async def renew() -> bool:
+        tries["renew"] += 1
+        return True
+
+    monkeypatch.setattr(workers, "register", register)
+    monkeypatch.setattr(svc, "renew", renew)
+    old = identity.owner()
+    fence.trip("测试")
+    svc._on_fenced()
+    assert identity.owner() != old and svc.pending_since is not None
+    task = asyncio.create_task(svc._lease_loop())
+    try:
+        for _ in range(300):
+            if tries["renew"] >= 2:
+                break
+            await asyncio.sleep(0.01)
+        assert not task.done(), "续租的任务还在"
+        assert tries["register"] == 3 and svc.pending_since is None and not fence.pending()
+        assert tries["renew"] >= 2, "登记上之后照常续租"
+    finally:
+        task.cancel()
+        with contextlib.suppress(asyncio.CancelledError):
+            await task