diff --git a/docs/CHANGELOG.md b/docs/CHANGELOG.md
index a48e4d955..de03a69b5 100644
--- a/docs/CHANGELOG.md
+++ b/docs/CHANGELOG.md
@@ -18,6 +18,29 @@ entry. See `CONTRIBUTING.md` § Releases & changelog.
 
 ## [Unreleased]
 
+### Fixed — the resume/reaper conformance test raced the wall clock
+
+- **`a RESUMED task survives the reaper` could fail for reasons unrelated to
+  the behaviour it guards.** It aged a parked task by `sleep(120)`, resumed it,
+  and then swept with `staleAfterMs: 60` using the reaper's *default real-time*
+  `now`. That left a 60ms budget between `provideInput` returning and the sweep
+  running: any stall longer than that — a GC pause, a loaded CI box, the serial
+  Postgres job sharing a runner — aged the freshly reset heartbeat past its own
+  window, so the reaper failed the task and the test went red while the code
+  was correct. Reproduced deterministically by inserting an 80ms stall before
+  the sweep.
+- The sweep now takes an explicit `now` anchored on the resume's own
+  `updatedAt`. `provideInput` writes `updatedAt` and `lastHeartbeatAt` in one
+  statement, so that timestamp comes from the same clock that stamped the row —
+  the database's for the durable store, the process's for the in-memory one —
+  and the comparison no longer depends on how long the test itself takes. With
+  the fix the test survives even a 3s injected stall.
+- `updatedAt` is the anchor precisely because it stays correct when the
+  behaviour regresses: dropping the `lastHeartbeatAt` reset still advances
+  `updatedAt`, so the frozen heartbeat lands outside the window and the test
+  goes red. Verified by mutation against **both** stores — in-memory and real
+  Postgres.
+
 ### Fixed — the rest of the test servers now bind the port they dial (#707)
 
 - **The remaining 57 `listen(0)` sites are converted.** #703 fixed the
diff --git a/middleware/test/tasks/taskStoreConformance.ts b/middleware/test/tasks/taskStoreConformance.ts
index 4b1220cb5..d412dba6a 100644
--- a/middleware/test/tasks/taskStoreConformance.ts
+++ b/middleware/test/tasks/taskStoreConformance.ts
@@ -292,10 +292,27 @@ export function runTaskStoreConformance(
         'provideInput advanced lastHeartbeatAt past creation',
       );
 
-      // Reap immediately with a window (60ms) that sits BETWEEN the frozen
-      // pre-park heartbeat (~120ms old) and the just-reset one (~0ms old). Uses
-      // the reaper's default real-time `now`, so both stores' wall clocks agree.
+      // Reap with a window (60ms) that sits BETWEEN the frozen pre-park
+      // heartbeat (>=120ms before the resume) and the just-reset one.
+      //
+      // `now` is anchored on the resume's OWN timestamp rather than the wall
+      // clock. `provideInput` writes `updatedAt` and `lastHeartbeatAt` in one
+      // statement, so `updatedAt` marks the instant of the resume in whichever
+      // clock the store uses — the database's for the durable store, the
+      // process's for the in-memory one — which is the same clock the row's
+      // `lastHeartbeatAt` was stamped by. Reading the default real-time `now`
+      // instead left a 60ms budget between `provideInput` returning and the
+      // sweep running: a GC pause or a loaded CI box aged the freshly reset
+      // heartbeat past its own window and failed this test for reasons that had
+      // nothing to do with the behaviour under test.
+      //
+      // `updatedAt` is the right anchor precisely because it stays correct when
+      // the behaviour under test regresses: drop the `lastHeartbeatAt` reset and
+      // `updatedAt` still advances, so the frozen heartbeat lands outside the
+      // window and this test goes red — which is the whole point of it.
+      const resumeAt = Date.parse(afterResume.updatedAt);
       const result = await store.reapOrphans({
+        now: new Date(resumeAt + 30),
         staleAfterMs: 60,
         purgeTerminalAfterMs: HOUR_MS,
       });