Context
On #19767 (4fed2f7c3a), two Brain-gated e2e shards failed both their attempt and their retry, then passed on re-run (run 38777334496, attempt 3). Sophie's receipt names the jobs (…827, …725); the shards' only code change was a unit fixture. Both failures come from the witness's own instrument, not from a product regression.
The Problem
WorkstationTabDragLabelOverlapNL.spec.mjs:57. The positive control "phase A exercised the sort machinery" asserts that a sampled frame shows the designed same-row slide-past crossing (sameLayer.length > 0). Locally on dev 7f0856c106 the spec failed 2 of 3 runs, always with 0 crossings, never a few. CI showed the same signature: canary 1/17/1, two DockFlip.play calls with geometryOnly: false. A probe of the same gesture recorded the strip's label order before and after. In 4 of 4 runs the order flipped, from ['Metrics', 'Audit'] to ['Audit', 'Metrics']; one of those runs sampled 0 crossings and the other three 10. The sort always runs. Whether a sampled frame catches the slide-past overlap is a race, so the control fails on a sort that happened.
WorkstationEdgeSplitterNL.spec.mjs:35. findEdgeSplitter lists the splitters, then calls getResizeConfig on each inside try/catch. A projection tick can replace a splitter generation between the two calls. The call then hits a destroyed id: the spec catches the throw, but Neo.ai.Client logs Failed to handle message Instance not found: neo-dashboard-dock-splitter-1 (src/ai/Client.mjs:295). The workerErrors fixture (test/playwright/fixtures.mjs:136) fails every test whose page received a worker error the test did not name. So the spec's own probe trips its own oracle, more often on slower runners.
The Architectural Reality
test/playwright/fixtures.mjs workerErrors: a test names an error it causes with workerErrors.expect(RegExp). Precedent: DockCrossWindowWitnessNL.spec.mjs:25, CanvasGroupPainting.spec.mjs:200.- The tab-drag witness's other terminal controls stay as they are: the canary,
DockFlip.play with geometryOnly: false, and the branch classification. The same-row crossings stay recorded as diagnostics, as the spec already labels them.
The Fix
WorkstationTabDragLabelOverlapNL.spec.mjs: read the strip's header-label order before and after the gesture, and replace the sampled-crossing control with "the order flipped". Keep the crossings as a logged diagnostic.WorkstationEdgeSplitterNL.spec.mjs: take the workerErrors fixture, and when a probe of a candidate id throws, name that id's Instance not found line, and only that id's.
Acceptance Criteria
Out of Scope
Whether the slide-past animation should always paint an overlapping frame, which is cosmetic and not this witness's defect. Neo.ai.Client logging failed requests it also returns to the caller is a product contract, left as it is.
Related
#19655 (the hosted tier, parent) · #19767 (where both surfaced) · #16406 (the tab-drag witness's defect) · #18705, #18521 (the last fixes to each spec)
Live latest-open sweep: latest 20 open issues at 2026-10-15T18:00:49Z; no equivalent found
A2A claim sweep: Sophie's 17:37Z receipt and my 17:41Z reply taking both; no other claim
MC sweep: "tab drag label overlap witness positive control same-row slide-past crossing flaky; edge splitter Instance not found dock splitter worker error", 5 results, no prior decision on either
Own-assignment sweep: 4 open (#19766, #19765, #19652, #15000), none overlapping
Retrieval Hint: "Workstation NL witness instrument flake order flip workerErrors probe miss"
Origin Session ID: 77c3c004-7908-4549-9190-b4f9c3cfa9dc
Context
On #19767 (
4fed2f7c3a), two Brain-gated e2e shards failed both their attempt and their retry, then passed on re-run (run 38777334496, attempt 3). Sophie's receipt names the jobs (…827, …725); the shards' only code change was a unit fixture. Both failures come from the witness's own instrument, not from a product regression.The Problem
WorkstationTabDragLabelOverlapNL.spec.mjs:57. The positive control "phase A exercised the sort machinery" asserts that a sampled frame shows the designed same-row slide-past crossing (sameLayer.length > 0). Locally ondev7f0856c106the spec failed 2 of 3 runs, always with 0 crossings, never a few. CI showed the same signature: canary 1/17/1, twoDockFlip.playcalls withgeometryOnly: false. A probe of the same gesture recorded the strip's label order before and after. In 4 of 4 runs the order flipped, from['Metrics', 'Audit']to['Audit', 'Metrics']; one of those runs sampled 0 crossings and the other three 10. The sort always runs. Whether a sampled frame catches the slide-past overlap is a race, so the control fails on a sort that happened.WorkstationEdgeSplitterNL.spec.mjs:35.findEdgeSplitterlists the splitters, then callsgetResizeConfigon each insidetry/catch. A projection tick can replace a splitter generation between the two calls. The call then hits a destroyed id: the spec catches the throw, butNeo.ai.ClientlogsFailed to handle message Instance not found: neo-dashboard-dock-splitter-1(src/ai/Client.mjs:295). TheworkerErrorsfixture (test/playwright/fixtures.mjs:136) fails every test whose page received a worker error the test did not name. So the spec's own probe trips its own oracle, more often on slower runners.The Architectural Reality
test/playwright/fixtures.mjsworkerErrors: a test names an error it causes withworkerErrors.expect(RegExp). Precedent:DockCrossWindowWitnessNL.spec.mjs:25,CanvasGroupPainting.spec.mjs:200.DockFlip.playwithgeometryOnly: false, and the branch classification. The same-row crossings stay recorded as diagnostics, as the spec already labels them.The Fix
WorkstationTabDragLabelOverlapNL.spec.mjs: read the strip's header-label order before and after the gesture, and replace the sampled-crossing control with "the order flipped". Keep the crossings as a logged diagnostic.WorkstationEdgeSplitterNL.spec.mjs: take theworkerErrorsfixture, and when a probe of a candidate id throws, name that id'sInstance not foundline, and only that id's.Acceptance Criteria
Instance not foundlines of ids its own probe hit; any other worker error still fails it.--repeat-each=4 --retries=0), with the receipt in the PR.Out of Scope
Whether the slide-past animation should always paint an overlapping frame, which is cosmetic and not this witness's defect.
Neo.ai.Clientlogging failed requests it also returns to the caller is a product contract, left as it is.Related
#19655 (the hosted tier, parent) · #19767 (where both surfaced) · #16406 (the tab-drag witness's defect) · #18705, #18521 (the last fixes to each spec)
Live latest-open sweep: latest 20 open issues at 2026-10-15T18:00:49Z; no equivalent found A2A claim sweep: Sophie's 17:37Z receipt and my 17:41Z reply taking both; no other claim MC sweep: "tab drag label overlap witness positive control same-row slide-past crossing flaky; edge splitter Instance not found dock splitter worker error", 5 results, no prior decision on either Own-assignment sweep: 4 open (#19766, #19765, #19652, #15000), none overlapping
Retrieval Hint: "Workstation NL witness instrument flake order flip workerErrors probe miss" Origin Session ID: 77c3c004-7908-4549-9190-b4f9c3cfa9dc