Context
test/playwright/e2e/dashboard/DockMotionNL.spec.mjs › a committed moveItem is FLIP-bracketed: the pane travels between real positions around the structural commit fails intermittently in the hosted e2e-brain tier, always with the same message:
the moved pane must land at a different position (from 1,70 to 1,70)
| Run |
Branch, head |
In-job retry |
39876017299 |
agent/20026-departed-window-last-resort, 654edbceac |
passed |
39963349591 |
agent/20056-feed-keeps-scroll, df39847e22 |
passed |
40007469729 |
push to dev, da49631cf6 |
failed too |
40046025369 |
PR #20087, 788d84e26e |
failed too; a job re-run passed |
These are at least four sightings in the last 120 runs of "E2E (Brain-gated tier)", back to 2026-10-22 14:56Z. I scanned the failed runs and the earlier attempts of re-run runs. A job whose in-job retry passed and stayed green would not show up, so this is a lower bound. The earliest sighting is 2026-10-23 00:06Z.
In every sighting, each assertion before this one passed: the operation applied, the document holds the item in its target node, and the witness recorded more than five frames, each with a non-zero width.
The Problem
- Cost: each red takes a diagnosis before someone re-runs it. On PR #20087 it held the reviewer seat until the re-run, and on
dev it reds the push. Under #19655's terminal predicate, a red in this tier blocks merge readiness, so a flake here blocks PRs that are fine.
- Not order-bound: the in-job retry runs in a fresh worker without the file's earlier tests, and it failed in two of the four sightings. The sightings also sat in different shard lineups, before and after the duration shards (
0ab667195e).
- No local repro so far: on macOS at
788d84e26e, the spec passed 7/7 alone and 9/9 in its CI shard's order.
The Architectural Reality
- The witness: every frame,
startMotionSampler records the first element carrying dock-flip-item-<id> (document.querySelector). Workspace#flipMarkerPrefix is stamped on every plain pane config, and the FLIP addon correlates panes by that class.
- The bracket: this test calls
expectSignalBracket(page) with the default SIGNAL, which matches any .neo-dashboard-dock-animating element. Several producers raise that class:
- the Workspace's FLIP play (
MotionSignal.enter(me) before flip.play, MotionSignal.leave(me) when it settles);
TabEnterButton;
RevealOverlay;
Maximize;
- the strip overflow projection (
LayoutAdapter's overflowProjectionStart / overflowProjectionIdle).
- Prior art in this file: #19893 diagnosed the same shape on the rail-click reveal SLIDES test. The any-element bracket opened and closed on another producer's signal, and
readMotionSamples stopped the witness before the motion under test was observed. PR #19894 gave that test a narrowed selector and a settle. The moveItem test still brackets any element and arms its witness without a settle.
- Design authority: the spec's docblock says committed dock operations "move real pixels", and that the signal "brackets each motion".
Diagnosis (AC-1, recorded 2026-10-26)
H2 holds. The evidence is six local failures under an 8× CDP CPU throttle (Emulation.setCPUThrottlingRate, see the repro comment): two on dev 622840cce2 and four on 07da0dac73. Their traces replace the two CI artifacts in step 1 of The Fix, so nothing was downloaded.
- The signal timeline is identical in all six. The bracket's appear-poll first saw 3 or 4 elements carrying
.neo-dashboard-dock-animating. Every one of them was a tab-header toolbar (neo-tab-header-toolbar-1 … -4). The next poll, under 25 ms later, saw none, so the bracket closed. The witness stopped within 10 ms after that.
- The Workspace's own signal came later. In one trace, the snapshot taken right after the witness stopped shows
neo-dock-workspace-1 carrying the signal. The FLIP play began after the bracket had already closed on the toolbars.
- The marker at the failing assertion is unique. There were 10
dock-flip-item-* elements carrying 10 distinct tokens, so no stale duplicate exists, and H1 is excluded. Every sample of the moved pane's marker read x: 1, y: 70, w: 770, its position before the move.
- H3 is not supported. In the one trace that ran long enough to show it, the Workspace began its FLIP play, which means the projection did move the pane.
The toolbars' signal comes from the same operation, because the move changes two tab strips. Settling earlier motion before arming the witness therefore cannot separate it. The fix narrows the bracket to the Workspace's own signal (.neo-dock-workspace plus .neo-dashboard-dock-animating), as #19894 did for the reveal overlay.
The Fix
- Read the recordings before fixing. Two artifacts hold
trace.zip and error-context.md for both failed attempts: test-results-e2e-brain-40046025369-shard-5 (177 MB) and test-results-e2e-brain-40007469729-shard-2 (73 MB). Both expire 2027-01-21/22. The trace shows when each producer's signal entered and left, against the moment the moved pane's DOM changed. Each hypothesis below needs that V-B-A:
- H2 (prior art #19893): the bracket closes on another producer's signal before the FLIP play raises its own, so the witness stops while the pane is still at its old position. This is test-only: bracket the Workspace's own FLIP signal, and settle earlier motion before arming the witness, as #19894 did for the reveal.
- H1: two elements carry the marker, and a stale one at the old position comes first in DOM order. Whether that is a product defect (a leftover pane) or a witness defect depends on what the stale element is.
- H3: the projection never moved the pane on that runner. This would be a product defect in the dock projection.
- Fix the owner the recordings convict, red-first on the recorded condition. If no recording separates the hypotheses, reproduce under CPU throttling first, since slow runners are the suspected condition.
Acceptance Criteria
Out of Scope
- Raising the retry count, quarantining the test, or loosening the assertion.
BROWSER JS ERROR: PlaywrightError: Target worker 'cw-win-a' does not exist. after DemoATourNL in the same shard. It prints in passing runs too (40039702013).
- The resizeSplit GLIDES test's any-element bracket (#19893's own out-of-scope row), unless the recordings show it shares the cause.
Related
- Parent: #19655 (the hosted Brain-gated tier).
- #19893 / PR #19894: the same witness shape on this file's reveal test.
- #14989: per-item markers for the geometry witnesses.
- PR #20087: a sighting (its re-run is green).
Live latest-open sweep: checked the latest 20 open issues at 2026-10-24T01:11:52Z and again at 01:14:43Z; no equivalent.
A2A in-flight claim sweep: the last 30 messages (all states) at 01:11:52Z and the last 8 at 01:14:43Z; no competing claim.
Exact sweep: gh search issues "DockMotionNL", "must land at a different position" and "moveItem FLIP", open and closed; no equivalent. Prior art: #19893, #14989, #15147, #14973.
MC sweep: "dock moveItem e2e the moved pane stays at its old position, flaky Brain-gated shard, retry fails, re-run passes" and "DockMotionNL moveItem FLIP-bracketed flake …"; 10 results each, no prior decision found.
Own-assignment sweep: 10 open; #19655 (the parent) overlaps by surface and parks no flake row.
Origin Session ID: 8ce18eb3-6705-48b9-ad43-b746ebce25b4
Retrieval Hint: "DockMotionNL moveItem FLIP moved pane from 1,70 to 1,70 any-element signal bracket"
Context
test/playwright/e2e/dashboard/DockMotionNL.spec.mjs› a committed moveItem is FLIP-bracketed: the pane travels between real positions around the structural commit fails intermittently in the hosted e2e-brain tier, always with the same message:39876017299agent/20026-departed-window-last-resort,654edbceac39963349591agent/20056-feed-keeps-scroll,df39847e2240007469729dev,da49631cf640046025369788d84e26eThese are at least four sightings in the last 120 runs of "E2E (Brain-gated tier)", back to 2026-10-22 14:56Z. I scanned the failed runs and the earlier attempts of re-run runs. A job whose in-job retry passed and stayed green would not show up, so this is a lower bound. The earliest sighting is 2026-10-23 00:06Z.
In every sighting, each assertion before this one passed: the operation applied, the document holds the item in its target node, and the witness recorded more than five frames, each with a non-zero width.
The Problem
devit reds the push. Under #19655's terminal predicate, a red in this tier blocks merge readiness, so a flake here blocks PRs that are fine.0ab667195e).788d84e26e, the spec passed 7/7 alone and 9/9 in its CI shard's order.The Architectural Reality
startMotionSamplerrecords the first element carryingdock-flip-item-<id>(document.querySelector).Workspace#flipMarkerPrefixis stamped on every plain pane config, and the FLIP addon correlates panes by that class.expectSignalBracket(page)with the defaultSIGNAL, which matches any.neo-dashboard-dock-animatingelement. Several producers raise that class:MotionSignal.enter(me)beforeflip.play,MotionSignal.leave(me)when it settles);TabEnterButton;RevealOverlay;Maximize;LayoutAdapter'soverflowProjectionStart/overflowProjectionIdle).readMotionSamplesstopped the witness before the motion under test was observed. PR #19894 gave that test a narrowedselectorand a settle. The moveItem test still brackets any element and arms its witness without a settle.Diagnosis (AC-1, recorded 2026-10-26)
H2 holds. The evidence is six local failures under an 8× CDP CPU throttle (
Emulation.setCPUThrottlingRate, see the repro comment): two on dev622840cce2and four on07da0dac73. Their traces replace the two CI artifacts in step 1 of The Fix, so nothing was downloaded..neo-dashboard-dock-animating. Every one of them was a tab-header toolbar (neo-tab-header-toolbar-1…-4). The next poll, under 25 ms later, saw none, so the bracket closed. The witness stopped within 10 ms after that.neo-dock-workspace-1carrying the signal. The FLIP play began after the bracket had already closed on the toolbars.dock-flip-item-*elements carrying 10 distinct tokens, so no stale duplicate exists, and H1 is excluded. Every sample of the moved pane's marker readx: 1, y: 70, w: 770, its position before the move.The toolbars' signal comes from the same operation, because the move changes two tab strips. Settling earlier motion before arming the witness therefore cannot separate it. The fix narrows the bracket to the Workspace's own signal (
.neo-dock-workspaceplus.neo-dashboard-dock-animating), as #19894 did for the reveal overlay.The Fix
trace.zipanderror-context.mdfor both failed attempts:test-results-e2e-brain-40046025369-shard-5(177 MB) andtest-results-e2e-brain-40007469729-shard-2(73 MB). Both expire 2027-01-21/22. The trace shows when each producer's signal entered and left, against the moment the moved pane's DOM changed. Each hypothesis below needs that V-B-A:Acceptance Criteria
Out of Scope
BROWSER JS ERROR: PlaywrightError: Target worker 'cw-win-a' does not exist.after DemoATourNL in the same shard. It prints in passing runs too (40039702013).Related
Live latest-open sweep: checked the latest 20 open issues at 2026-10-24T01:11:52Z and again at 01:14:43Z; no equivalent. A2A in-flight claim sweep: the last 30 messages (all states) at 01:11:52Z and the last 8 at 01:14:43Z; no competing claim. Exact sweep:
gh search issues "DockMotionNL","must land at a different position"and"moveItem FLIP", open and closed; no equivalent. Prior art: #19893, #14989, #15147, #14973. MC sweep: "dock moveItem e2e the moved pane stays at its old position, flaky Brain-gated shard, retry fails, re-run passes" and "DockMotionNL moveItem FLIP-bracketed flake …"; 10 results each, no prior decision found. Own-assignment sweep: 10 open; #19655 (the parent) overlaps by surface and parks no flake row.Origin Session ID: 8ce18eb3-6705-48b9-ad43-b746ebce25b4 Retrieval Hint: "DockMotionNL moveItem FLIP moved pane from 1,70 to 1,70 any-element signal bracket"