diff --git a/backend/app/cluster/slots.py b/backend/app/cluster/slots.py
index a7c83136..d363dedb 100644
--- a/backend/app/cluster/slots.py
+++ b/backend/app/cluster/slots.py
@@ -194,6 +194,8 @@ class _Holder:
     run_id: str
     loop: asyncio.AbstractEventLoop
     woke: asyncio.Event = field(default_factory=asyncio.Event)
+    #: 还在排队等名额（发下来之前）
+    queued: bool = True
 
 
 _holders: dict[str, _Holder] = {}
@@ -379,6 +381,7 @@ async def hold(run_id: str, origin_kind: str | None) -> AsyncIterator[str]:
             await grant()
             got = await _granted(run_id)
             if got:
+                holder.queued = False
                 break
             if got is None:
                 await _request(run_id, klass, priority)
@@ -397,6 +400,17 @@ async def hold(run_id: str, origin_kind: str | None) -> AsyncIterator[str]:
         signals.publish("slot")
 
 
+async def waiting(run_id: str, *, here: bool = False) -> bool:
+    """这次运行是不是正排着队等名额（评测每题的时限、算期望的时限不算这一段，eval/runner、eval/gold）。
+    here：这次运行的执行任务在本进程里（还没开始要、或者已经放掉了名额的就不是在等，不用查库）。"""
+    holder = _holders.get(run_id)
+    if holder is not None:
+        return holder.queued
+    if here:
+        return False
+    return (await _granted(run_id)) is False
+
+
 async def snapshot() -> dict[str, Any]:
     """现在各类在跑、在等几个（系统信息、验收脚本看）。"""
     if _in_memory():
@@ -413,4 +427,5 @@ async def snapshot() -> dict[str, Any]:
 
 
 __all__ = ["CLASSES", "CLASS_LABEL", "ORIGIN_CLASS", "SETTING_KEY", "borrowable", "class_of", "defaults", "forget_policy", "grant",
+           "waiting",
            "hold", "merged", "policy", "problem", "snapshot", "values", "wake_local"]
diff --git a/backend/app/core/job_runtime.py b/backend/app/core/job_runtime.py
index a9f103e6..b050ee83 100644
--- a/backend/app/core/job_runtime.py
+++ b/backend/app/core/job_runtime.py
@@ -241,16 +241,40 @@ class SourceLanes:
 
             await session.execute(text("SELECT pg_advisory_xact_lock(:k)"), {"k": advisory_key(f"lane:{source_id}")})
 
-    async def _rows(self, session: AsyncSession, source_id: str) -> list[Any]:
-        """这个源队里活着的各行（号小的在前）；属主不在了的删掉（不提交）。"""
-        from app.cluster import state
+    async def _dead_owners(self, source_id: str) -> frozenset[str]:
+        """拿这个源的锁之前备好：队里别的进程的属主哪些确实不在了（手上那份不认识的当场再查一次登记）。持着咨询锁时不另取连接
+        查库（复核：连接池用尽时会卡到取连接超时）；拿锁之后才出现的属主不在这里面，这一次当活着，下一次再说。"""
+        from app.cluster import identity, state
+        from app.db.models_cluster import LaneTicket as Row
+
+        async with SessionLocal() as session:
+            owners = set((await session.execute(select(Row.owner).where(Row.source_id == source_id).distinct()))
+                         .scalars())
+        me = identity.owner()
+        others = {o for o in owners if o and o != me}
+        if not others:
+            return frozenset()
+        live = await state.live_for(others)
+        return frozenset(o for o in others if o not in live.others)
+
+    async def _rows(self, session: AsyncSession, source_id: str, dead_owners: frozenset[str]) -> list[Any]:
+        """这个源队里活着的各行（号小的在前）；属主不在了的删掉（不提交）。dead_owners 是拿锁之前备好的那一份：属主是别的进程的
+        只认它；属主是自己的看本进程里登记着没有（正有人排进来时一律算活着）。"""
+        from app.cluster import identity
         from app.db.models_cluster import LaneTicket as Row
 
         rows = list((await session.execute(select(Row).where(Row.source_id == source_id).order_by(Row.id))).scalars())
-        live = await state.live_for([r.owner for r in rows])
-        if self._entering.get(source_id):
-            live = live.trusting()
-        dead = [r.id for r in rows if not live.alive(r.owner, local=self._local_alive(r.id))]
+        me = identity.owner()
+        entering = bool(self._entering.get(source_id))
+
+        def alive(row: Any) -> bool:
+            if not row.owner:
+                return False
+            if row.owner == me:
+                return entering or self._local_alive(row.id)
+            return row.owner not in dead_owners
+
+        dead = [r.id for r in rows if not alive(r)]
         if dead:
             from sqlalchemy import delete
 
@@ -262,9 +286,10 @@ class SourceLanes:
         from app.db.models_cluster import LaneTicket as Row
 
         granted_now: int | None = None
+        dead_owners = await self._dead_owners(source_id)
         async with self._lock(), SessionLocal() as session:
             await self._locked(session, source_id)
-            rows = await self._rows(session, source_id)
+            rows = await self._rows(session, source_id, dead_owners)
             head = next((r for r in rows if r.granted), None)
             if head is None and rows:
                 head = rows[0]
@@ -318,9 +343,10 @@ class SourceLanes:
         if kind not in _LANE_KINDS:
             raise ValueError(f"同源队里没有登记「{kind}」这种事，先 register_lane_kind")
         self._listen()
+        dead_owners = await self._dead_owners(source_id)
         async with self._lock(), SessionLocal() as session:
             await self._locked(session, source_id)
-            rows = await self._rows(session, source_id)
+            rows = await self._rows(session, source_id, dead_owners)
             if rows:
                 head = next((r for r in rows if r.granted), rows[0])
                 await session.commit()
@@ -337,8 +363,9 @@ class SourceLanes:
 
     async def holders(self, source_id: str) -> list[LaneHolder]:
         """队里的各位，排在最前面（正在做）的在前。"""
+        dead_owners = await self._dead_owners(source_id)
         async with SessionLocal() as session:
-            rows = await self._rows(session, source_id)
+            rows = await self._rows(session, source_id, dead_owners)
             await session.commit()
         rows.sort(key=lambda r: (not r.granted, r.id))
         return [LaneHolder(r.kind, r.job_id, r.owner) for r in rows]
diff --git a/backend/app/eval/answer.py b/backend/app/eval/answer.py
index 6acc25cf..2842b27e 100644
--- a/backend/app/eval/answer.py
+++ b/backend/app/eval/answer.py
@@ -24,7 +24,7 @@ app/chat/job.py），生成过程是 api/copilot.answer_ops，几个「决定」
 用量按角色记（做法 3、11）：助手 = 生成（answer_ops 末尾的 usage）+ 复核（review_of_run 的 usage）；写作 = 答案运行里的
 llm.end（按模型分）。重来的话两次都算——那都是这一题花的钱。
 
-时限、停止由调用方管（runner.py 用 asyncio.wait_for 包住 drive）：这里随时把正在跑的那次运行记在 current_run，调用方
+时限、停止由调用方管（runner.py 用 within 包住 drive，排队等名额的那一段不算）：这里随时把正在跑的那次运行记在 current_run，调用方
 超时、停止时据此取消它（stop_run）。
 """
 from __future__ import annotations
diff --git a/backend/app/eval/gold.py b/backend/app/eval/gold.py
index 872aaf32..4b584f1b 100644
--- a/backend/app/eval/gold.py
+++ b/backend/app/eval/gold.py
@@ -209,11 +209,19 @@ async def _launch_and_wait(graph: dict[str, Any], payload: dict[str, Any], *, na
             run_id = run.id
     except LaunchRefused as e:
         return _Ran("failed", None, f"算期望的运行没有发起：{e.message}", GOLD_FAILED)
+    from app.cluster import slots
+
     loop = asyncio.get_running_loop()
     deadline = loop.time() + max(0.0, float(ctx.timeout_s))
+    last = loop.time()
     while True:
         # 执行任务收完尾（含封存）才回来；任务已经不在了（刚建好还没排上、或者已经结束）就立刻回来，再按库里的状态判
-        await run_manager.wait_idle(run_id, timeout=max(0.05, min(5.0, deadline - loop.time())))
+        await run_manager.wait_idle(run_id, timeout=max(0.05, min(1.0, deadline - loop.time())))
+        now = loop.time()
+        if await slots.waiting(run_id, here=run_manager.is_active(run_id)):
+            # 排队等名额的那一段不算时限（复核：满载时排队不该算进每题时限）
+            deadline += now - last
+        last = now
         async with SessionLocal() as session:
             row = await session.get(Run, run_id)
             status, error = (row.status, row.error) if row is not None else ("failed", "运行记录不见了")
diff --git a/backend/app/eval/runner.py b/backend/app/eval/runner.py
index 441430d7..af8718f7 100644
--- a/backend/app/eval/runner.py
+++ b/backend/app/eval/runner.py
@@ -15,7 +15,7 @@ snapshot 视图）：起草、剖析照常进行，也不影响这次评测的
 **并发。** 同时最多 max_parallel 题，且所有臂合起来不超过全局运行并发的四分之一。评测发起的作答、算期望的运行归「系统」类
 （app/cluster/slots.py）：全站的名额按类别有上限和优先级，评测再多也只占得到系统那一类的名额，问数据、试跑优先，不被挤占。
 A/B 的同一题几臂同时开跑。每题时限 per_case_timeout_s：作答一臂超时就取消它正在跑的运行、记超时；
-期望探针每次也按这个时限。
+期望探针每次也按这个时限。排队等运行名额的那一段不算时限（within、gold._launch_and_wait）：满载时排队不该被记成超时、触发止损。
 
 **预算**（评测-D9；单独一本账 engine/spend.eval_spend，不挤占裁判、核查的额度）：开跑前按上一次的每题平均花费估算（estimate）。
 每题开评前按估算预留——加上会超过这次评测的金额上限、或评测每日上限就不再开新题，剩下的记「已到上限」；评完按实际找补。
@@ -33,6 +33,7 @@ A/B 的同一题几臂同时开跑。每题时限 per_case_timeout_s：作答一
 from __future__ import annotations
 
 import asyncio
+import contextlib
 import copy
 import logging
 from collections.abc import Mapping, Sequence
@@ -519,9 +520,9 @@ class EvalExecution(runtime.JobRun):
         try:
             if case.get("kind") == "workflow":
                 graph, inputs = await self._workflow_graph(case, values)
-                await asyncio.wait_for(driver.drive_workflow(graph, inputs), timeout=timeout)
+                await within(driver.drive_workflow(graph, inputs), timeout, lambda: _queued(driver))
             else:
-                await asyncio.wait_for(driver.drive(), timeout=timeout)
+                await within(driver.drive(), timeout, lambda: _queued(driver))
         except asyncio.TimeoutError:
             turn = driver.turn
             turn.timed_out = True
@@ -964,6 +965,40 @@ def running(run: Any) -> bool:
     return manager.running(leader_of(run))
 
 
+async def _queued(driver: Any) -> bool:
+    """这一题的作答运行正排着队等名额（app/cluster/slots）：这一段不算每题的时限。"""
+    from app.cluster import slots
+    from app.engine.runner import run_manager
+
+    run_id = getattr(driver.turn, "current_run", None)
+    return bool(run_id) and await slots.waiting(run_id, here=run_manager.is_active(run_id))
+
+
+async def within(coro: Any, budget: float, paused: Any, *, tick: float = 0.5) -> Any:
+    """同 asyncio.wait_for(coro, budget)，只是 paused() 为真的那几段不算时间（复核：满载时排队等名额不该算进每题时限、触发止损）。
+    超时取消 coro、抛 TimeoutError。"""
+    loop = asyncio.get_running_loop()
+    task = asyncio.ensure_future(coro)
+    spent, last = 0.0, loop.time()
+    try:
+        while True:
+            done, _ = await asyncio.wait({task}, timeout=max(0.02, min(tick, budget - spent)))
+            if task in done:
+                return task.result()
+            now = loop.time()
+            step, last = now - last, now
+            if not await paused():
+                spent += step
+            if spent >= budget:
+                task.cancel()
+                with contextlib.suppress(asyncio.CancelledError, Exception):
+                    await task
+                raise asyncio.TimeoutError()
+    except asyncio.CancelledError:
+        task.cancel()
+        raise
+
+
 def leader_of_id(rows: Sequence[EvalRun], job_id: str) -> str:
     """rows 里 id 是 job_id 的那一臂由哪一臂的任务在跑（A/B 几臂共用领头那一臂的后台任务）。"""
     row = next((r for r in rows if r.id == job_id), None)
diff --git a/backend/tests/test_cluster.py b/backend/tests/test_cluster.py
index 2243dc6f..58c6564a 100644
--- a/backend/tests/test_cluster.py
+++ b/backend/tests/test_cluster.py
@@ -230,6 +230,54 @@ async def test_the_source_lane_is_shared_and_skips_tickets_of_a_gone_process():
     await other.release()
 
 
+async def test_the_source_lane_does_not_look_up_liveness_while_holding_its_lock(monkeypatch):
+    """复核：持着这个源的咨询锁时不另取连接查「谁还活着」（连接池用尽时会卡到取连接超时）——拿锁之前备好；拿锁之后才出现的
+    别的属主这一次当活着，不删它的行。"""
+    from app.core import job_runtime as runtime
+
+    import app.data.catalog_jobs  # noqa: F401 - 登记 draft、profile
+
+    lanes = runtime.lanes
+    holding = {"now": False, "looked_inside": 0}
+    real_locked, real_rows, real_live_for = lanes._locked, lanes._rows, state.live_for
+
+    async def locked(session, source_id):
+        await real_locked(session, source_id)
+        holding["now"] = True
+
+    async def rows(session, source_id, dead_owners):
+        try:
+            return await real_rows(session, source_id, dead_owners)
+        finally:
+            holding["now"] = False
+
+    async def live_for(owners):
+        if holding["now"]:
+            holding["looked_inside"] += 1
+        return await real_live_for(owners)
+
+    monkeypatch.setattr(lanes, "_locked", locked)
+    monkeypatch.setattr(lanes, "_rows", rows)
+    monkeypatch.setattr(state, "live_for", live_for)
+    async with SessionLocal() as session:
+        session.add(LaneTicket(source_id="src-lock", kind="profile", job_id="ghost-job", owner=GHOST, granted=True))
+        await session.commit()
+    mine = await lanes.enter("src-lock", "draft", "job-lock")
+    assert mine.granted.is_set(), "没了的进程占着的位置照旧让出来（拿锁之前就查清楚了）"
+    assert holding["looked_inside"] == 0
+    await mine.release()
+
+    async with SessionLocal() as session:
+        session.add(LaneTicket(source_id="src-late", kind="profile", job_id="late-job", owner=GHOST, granted=True))
+        await session.commit()
+        kept = await real_rows(session, "src-late", frozenset())
+        await session.commit()
+    assert [r.owner for r in kept] == [GHOST], "拿锁之前那份里没有的属主这一次当活着"
+    async with SessionLocal() as session:
+        await session.execute(delete(LaneTicket).where(LaneTicket.source_id.in_(["src-lock", "src-late"])))
+        await session.commit()
+
+
 async def test_reaping_runs_only_touches_the_ones_whose_owner_is_gone():
     from app.engine.runner import run_manager
 
diff --git a/backend/tests/test_eval_runner.py b/backend/tests/test_eval_runner.py
index a9685e62..94a9c106 100644
--- a/backend/tests/test_eval_runner.py
+++ b/backend/tests/test_eval_runner.py
@@ -289,6 +289,87 @@ async def test_a_case_over_its_time_limit_stops_its_run(monkeypatch):
     assert item.status == "ok" and item.path == "build_failed" and "每题时限" in item.error
 
 
+async def test_time_queued_for_a_run_slot_does_not_count_against_the_case_limit(monkeypatch):
+    """复核：满载时作答运行排队等名额的那一段不算每题时限（排了 0.8 秒、时限 0.3 秒，拿到名额后很快答完：不记超时）。"""
+    from app.cluster import slots
+
+    stored = slots.values({"max_total": 2, "interactive_reserve": 1})
+
+    async def fixed(*, fresh: bool = False):  # noqa: ARG001
+        return stored
+
+    monkeypatch.setattr(slots, "policy", fixed)
+    slots.forget_policy()
+    gate = asyncio.Event()
+
+    async def blocker() -> None:
+        async with slots.hold("eval-slot-blocker", "eval"):
+            await gate.wait()
+
+    async def drive(self):
+        self.turn.current_run = "eval-answer-in-queue"
+        await self._phase("build")
+        async with slots.hold(self.turn.current_run, "eval"):
+            await asyncio.sleep(0.05)
+        self.turn.path = "asked"
+        self.turn.usage["assistant"] = {"by_model": {}, "input_tokens": 0, "output_tokens": 0, "calls": 1,
+                                        "cost_usd": 0.0, "unpriced_calls": 0}
+        return self.turn
+
+    monkeypatch.setattr(answer_mod.Driver, "drive", drive)
+    original = runner.EvalExecution._prepare
+
+    async def prepare(self):
+        units = await original(self)
+        self.budget["per_case_timeout_s"] = 0.3
+        return units
+
+    monkeypatch.setattr(runner.EvalExecution, "_prepare", prepare)
+    held = asyncio.create_task(blocker())
+    try:
+        while not (await slots.snapshot())["classes"]["system"]["running"]:
+            await asyncio.sleep(0.02)
+        suite_id = await es.make_suite("none", cases=ask_cases(1))
+        [run] = await start(suite_id)
+        while not await slots.waiting("eval-answer-in-queue"):
+            await asyncio.sleep(0.02)
+        await asyncio.sleep(0.8)
+        gate.set()
+        await es.wait_eval(run.id)
+    finally:
+        gate.set()
+        await held
+    [item] = await results(run.id)
+    assert item.status == "ok" and item.path == "asked" and not item.error, item.error
+
+
+async def test_within_does_not_count_the_paused_stretches():
+    paused = {"now": True}
+
+    async def is_paused() -> bool:
+        return paused["now"]
+
+    async def work() -> str:
+        await asyncio.sleep(0.5)
+        paused["now"] = False
+        await asyncio.sleep(0.05)
+        return "done"
+
+    assert await runner.within(work(), 0.2, is_paused, tick=0.02) == "done", "暂停的那 0.5 秒不算"
+
+    async def never():
+        await asyncio.sleep(10)
+
+    async def awake() -> bool:
+        return False
+
+    loop = asyncio.get_running_loop()
+    t0 = loop.time()
+    with pytest.raises(asyncio.TimeoutError):
+        await runner.within(never(), 0.2, awake, tick=0.02)
+    assert loop.time() - t0 < 1.0
+
+
 async def test_a_run_stopped_at_an_approval_is_cancelled_not_approved(client, monkeypatch, tmp_path):
     from app.eval.answer import APPROVAL_REASON