Context
PR #18174 (Resolves #18173) went red on unit in CI at run 33737551415 while passing every local arm. @neo-opus-ada independently armed DragCoordinator.spec.mjs 3085 times locally without a single reproduction and reported it as "a real CI-only unit failure I cannot reproduce".
It is not reproducible from that file because the defect is not in it. test/playwright/unit/draggable/dashboard/SortZone.spec.mjs mutates the shared Neo.manager.DragCoordinator singleton and never restores the engine value.
Live latest-open sweep: checked the latest 20 open issues at 2026-09-03T09:31:02Z; no equivalent found. Closed-queue sweep on nativeWindowDropDwellMs returns only #18115 (the feature, closed). Own-assignment sweep: no same-surface hit. Memory Core rationale sweep surfaced the class as recurring but not this instance — see Related.
The Problem
test/playwright/unit/draggable/dashboard/SortZone.spec.mjs:60 runs an afterEach whose stated purpose is to undo the suite's mutations. It is careful and well-commented about the five stubbed methods below it — its own comment says "a stub that survives its suite is indistinguishable from the method simply not working". Two lines above those deletions it does this:
DragCoordinator.nativeWindowDropDwellMs = 450;
DragCoordinator.nativeWindowDropSettleMs = 250;
No original is captured anywhere in the file. These are the values this fixture wants, written in the place where the engine's values belong. The engine defaults are src/manager/DragCoordinator.mjs:31 nativeWindowDropDwellMs: 1200 and src/manager/DragCoordinator.mjs:89 nativeWindowDropSettleMs: 250.
nativeWindowDropSettleMs happens to coincide with its default, so it is harmless today and harmful the moment the default moves. nativeWindowDropDwellMs does not coincide: after this suite runs, every later spec sharing the worker sees a dwell of 450 instead of 1200.
Whether DragCoordinator.spec.mjs lands in that worker after SortZone.spec.mjs is worker-assignment order. That is why it is CI-only, and why Playwright reports it as 1 flaky — it passes on the retry, because a retry gets a fresh worker, so a cross-file isolation leak heals by construction on every re-run. The job still exits 1.
Both CI numbers fall out of the leak exactly
The failing assertion is DragCoordinator.spec.mjs:817, expect(delays.at(-1)).toBe(DragCoordinator.nativeWindowDropDwellMs - 600).
| Reported |
Arithmetic |
Reading |
Expected: -150 |
450 - 600 |
The assertion re-reads the config at assertion time and gets the leaked 450, not 1200. A negative expectation is itself the tell: the test cannot mean to expect a negative remaining hold. |
Received: 250 |
Math.max(0, 450 - 600) → 0 |
With dwell at 450 the hold is already spent by frame 3, so DragCoordinator.mjs:1234 clamps the remainder to 0 and the gesture completes mid-test. The commit then schedules the settle timer, so delays.at(-1) is the 250ms settle interval, not a dwell remainder. |
At the true default of 1200 the same three frames leave 1200 - 600 = 600 remaining, the hold stays armed, no settle timer is scheduled, and the assertion holds. The test is sound; it is measuring a coordinator whose dwell another file quietly halved.
The Architectural Reality
Neo.manager.DragCoordinator is a singleton, resolved through the Neo.manager namespace and shared by every spec in a Playwright worker. There is no per-spec instance to isolate behind.
test/playwright/unit/draggable/dashboard/SortZone.spec.mjs:60-90 — the afterEach. Its stub-deletion block is correct (it deletes own properties to re-expose the prototype). The two config assignments above it are the defect: assignment where restoration was intended.
src/manager/DragCoordinator.mjs:31 / :89 — the declared defaults, and the only authority for what "restore" means.
src/manager/DragCoordinator.mjs:1234 — dwellRemaining = Math.max(0, me.nativeWindowDropDwellMs - (now - firstSeenAt)), the clamp that converts the leaked value into a completed gesture rather than an obviously wrong number.
test/playwright/unit/manager/DragCoordinator.spec.mjs:812 — the victim, not the culprit.
Note SortZone.spec.mjs:641 and :725 also set nativeWindowDropDwellMs: 0, but those are per-fixture config objects, not singleton writes, and are out of scope here.
The Fix
In test/playwright/unit/draggable/dashboard/SortZone.spec.mjs, capture the engine's own values once before the suite mutates them and restore those, so the restore cannot drift from the default:
- Read
nativeWindowDropDwellMs and nativeWindowDropSettleMs off the singleton in a beforeAll (or at first beforeEach) into module-scoped constants.
- In
afterEach, assign the captured values back instead of the literals 450 / 250.
The suite may still set the dwell to 450 while it runs — that is a legitimate fixture need. What it may not do is hand 450 to the next spec.
Acceptance Criteria
Out of Scope
- Any change to
DragCoordinator.spec.mjs or to the #18173 dwell-restart behaviour it covers. That test is correct as written.
- The
nativeWindowDropDwellMs: 0 fixture configs at SortZone.spec.mjs:641 / :725 — per-instance config, not singleton state.
- A general sweep for other singleton-leaking specs. Worth doing, and it is not this ticket; see #17922 for the adjacent coverage lane.
- Any change to how Playwright assigns specs to workers, or to
failOnFlakyTests.
Avoided Traps
- Retrying, quarantining, or
test.retry-ing DragCoordinator.spec.mjs:812. The industry-default response to a flaky label is to re-run or mark it flaky. Here the retry is the mask: a fresh worker heals the leak by construction, so "it passed on re-run" is the signature of an isolation defect, not evidence of runner variance. Quarantining would have discarded a correct test and left the leak to hit the next arrival.
- Hunting the defect inside
DragCoordinator.spec.mjs. 3085 local arms of that file could not have reproduced it; the file is the victim. The discriminating instrument is a two-file, single-worker ordered run — nothing scoped to one file can see a cross-file leak.
- "Fixing" the assertion to expect 450. It would go green and cement the leak as the contract, and the negative expectation (
-150) would have been left standing as if intended.
- Adding a defensive re-assert of the default at the top of
DragCoordinator.spec.mjs. That treats the symptom at the victim and leaves every other later spec in the worker exposed.
Decision Record impact
none — this is a test-isolation defect, not an architectural change. The engine defaults at src/manager/DragCoordinator.mjs:31/:89 remain the single source of authority; the fix makes the spec derive from them instead of restating them.
Related
- PR #18174 / #18173 — the blocked PR and lane where this surfaced. This is its only remaining CI blocker.
- #18115 — the closed feature that introduced the dwell/settle protocol.
- #12659 — prior instance of the same class:
QueryRecentTurns.spec.mjs stubbed the TextEmbeddingService.embedText singleton method without restoring it, reddening dev and every PR through order-dependence.
- #15789 / PR #15797 — a previous
SortZone bleed, same file, same "test that lies" shape.
- #17922 — the buildScripts specs lost in the split; adjacent test-coverage lane.
Handoff Retrieval Hints
query_raw_memories: "cross-file singleton leak heals on retry fresh worker" — the 2026-08-19 note establishing that a green same-SHA re-run is what a leak looks like, not what a flake looks like.
query_raw_memories: "SortZone bleed test that lies" — the #15789 precedent in the same file.
- Reproducer anchor: CI run
33737551415, job 100591868953, head 2cd822b842.
unowned-rationale: parked, not dropped. I diagnosed this from CI logs during a nightshift heartbeat; the working tree in this checkout is held by a concurrent @neo-opus-grace session, so I am not taking a second branch this run rather than risk racing it. The fix is two lines and fully specified above — any seat may claim it. If it is still unowned on my next pass, I take it.
Origin Session ID: d1438cf3-b279-4330-8b3d-bbff2fdca44e
Context
PR #18174 (
Resolves #18173) went red onunitin CI at run 33737551415 while passing every local arm. @neo-opus-ada independently armedDragCoordinator.spec.mjs3085 times locally without a single reproduction and reported it as "a real CI-only unit failure I cannot reproduce".It is not reproducible from that file because the defect is not in it.
test/playwright/unit/draggable/dashboard/SortZone.spec.mjsmutates the sharedNeo.manager.DragCoordinatorsingleton and never restores the engine value.Live latest-open sweep: checked the latest 20 open issues at 2026-09-03T09:31:02Z; no equivalent found. Closed-queue sweep on
nativeWindowDropDwellMsreturns only #18115 (the feature, closed). Own-assignment sweep: no same-surface hit. Memory Core rationale sweep surfaced the class as recurring but not this instance — see Related.The Problem
test/playwright/unit/draggable/dashboard/SortZone.spec.mjs:60runs anafterEachwhose stated purpose is to undo the suite's mutations. It is careful and well-commented about the five stubbed methods below it — its own comment says "a stub that survives its suite is indistinguishable from the method simply not working". Two lines above those deletions it does this:DragCoordinator.nativeWindowDropDwellMs = 450; DragCoordinator.nativeWindowDropSettleMs = 250;No original is captured anywhere in the file. These are the values this fixture wants, written in the place where the engine's values belong. The engine defaults are
src/manager/DragCoordinator.mjs:31nativeWindowDropDwellMs: 1200andsrc/manager/DragCoordinator.mjs:89nativeWindowDropSettleMs: 250.nativeWindowDropSettleMshappens to coincide with its default, so it is harmless today and harmful the moment the default moves.nativeWindowDropDwellMsdoes not coincide: after this suite runs, every later spec sharing the worker sees a dwell of 450 instead of 1200.Whether
DragCoordinator.spec.mjslands in that worker afterSortZone.spec.mjsis worker-assignment order. That is why it is CI-only, and why Playwright reports it as1 flaky— it passes on the retry, because a retry gets a fresh worker, so a cross-file isolation leak heals by construction on every re-run. The job still exits 1.Both CI numbers fall out of the leak exactly
The failing assertion is
DragCoordinator.spec.mjs:817,expect(delays.at(-1)).toBe(DragCoordinator.nativeWindowDropDwellMs - 600).Expected: -150450 - 600Received: 250Math.max(0, 450 - 600)→0DragCoordinator.mjs:1234clamps the remainder to0and the gesture completes mid-test. The commit then schedules the settle timer, sodelays.at(-1)is the 250ms settle interval, not a dwell remainder.At the true default of 1200 the same three frames leave
1200 - 600 = 600remaining, the hold stays armed, no settle timer is scheduled, and the assertion holds. The test is sound; it is measuring a coordinator whose dwell another file quietly halved.The Architectural Reality
Neo.manager.DragCoordinatoris a singleton, resolved through theNeo.managernamespace and shared by every spec in a Playwright worker. There is no per-spec instance to isolate behind.test/playwright/unit/draggable/dashboard/SortZone.spec.mjs:60-90— theafterEach. Its stub-deletion block is correct (itdeletes own properties to re-expose the prototype). The two config assignments above it are the defect: assignment where restoration was intended.src/manager/DragCoordinator.mjs:31/:89— the declared defaults, and the only authority for what "restore" means.src/manager/DragCoordinator.mjs:1234—dwellRemaining = Math.max(0, me.nativeWindowDropDwellMs - (now - firstSeenAt)), the clamp that converts the leaked value into a completed gesture rather than an obviously wrong number.test/playwright/unit/manager/DragCoordinator.spec.mjs:812— the victim, not the culprit.Note
SortZone.spec.mjs:641and:725also setnativeWindowDropDwellMs: 0, but those are per-fixture config objects, not singleton writes, and are out of scope here.The Fix
In
test/playwright/unit/draggable/dashboard/SortZone.spec.mjs, capture the engine's own values once before the suite mutates them and restore those, so the restore cannot drift from the default:nativeWindowDropDwellMsandnativeWindowDropSettleMsoff the singleton in abeforeAll(or at firstbeforeEach) into module-scoped constants.afterEach, assign the captured values back instead of the literals450/250.The suite may still set the dwell to 450 while it runs — that is a legitimate fixture need. What it may not do is hand 450 to the next spec.
Acceptance Criteria
SortZone.spec.mjscaptures the engine defaults rather than hardcoding450/250in itsafterEach.SortZone.spec.mjssuite completes,Neo.manager.DragCoordinator.nativeWindowDropDwellMs === 1200andnativeWindowDropSettleMs === 250.SortZone.spec.mjsandDragCoordinator.spec.mjsin that order in a single worker fails atDragCoordinator.spec.mjs:817withExpected: -150 / Received: 250; with the fix applied, the same ordered single-worker run passes. A green run alone does not satisfy this AC — the red arm must be demonstrated, and it must be red for this reason.nativeWindowDropDwellMsinsrc/manager/DragCoordinator.mjsdoes not require a corresponding edit inSortZone.spec.mjs.unitjob reaches green with zeroflakyentries.Out of Scope
DragCoordinator.spec.mjsor to the #18173 dwell-restart behaviour it covers. That test is correct as written.nativeWindowDropDwellMs: 0fixture configs atSortZone.spec.mjs:641/:725— per-instance config, not singleton state.failOnFlakyTests.Avoided Traps
test.retry-ingDragCoordinator.spec.mjs:812. The industry-default response to aflakylabel is to re-run or mark it flaky. Here the retry is the mask: a fresh worker heals the leak by construction, so "it passed on re-run" is the signature of an isolation defect, not evidence of runner variance. Quarantining would have discarded a correct test and left the leak to hit the next arrival.DragCoordinator.spec.mjs. 3085 local arms of that file could not have reproduced it; the file is the victim. The discriminating instrument is a two-file, single-worker ordered run — nothing scoped to one file can see a cross-file leak.-150) would have been left standing as if intended.DragCoordinator.spec.mjs. That treats the symptom at the victim and leaves every other later spec in the worker exposed.Decision Record impact
none— this is a test-isolation defect, not an architectural change. The engine defaults atsrc/manager/DragCoordinator.mjs:31/:89remain the single source of authority; the fix makes the spec derive from them instead of restating them.Related
QueryRecentTurns.spec.mjsstubbed theTextEmbeddingService.embedTextsingleton method without restoring it, reddeningdevand every PR through order-dependence.SortZonebleed, same file, same "test that lies" shape.Handoff Retrieval Hints
query_raw_memories:"cross-file singleton leak heals on retry fresh worker"— the 2026-08-19 note establishing that a green same-SHA re-run is what a leak looks like, not what a flake looks like.query_raw_memories:"SortZone bleed test that lies"— the #15789 precedent in the same file.33737551415, job100591868953, head2cd822b842.unowned-rationale: parked, not dropped. I diagnosed this from CI logs during a nightshift heartbeat; the working tree in this checkout is held by a concurrent @neo-opus-grace session, so I am not taking a second branch this run rather than risk racing it. The fix is two lines and fully specified above — any seat may claim it. If it is still unowned on my next pass, I take it.
Origin Session ID: d1438cf3-b279-4330-8b3d-bbff2fdca44e