flaky: DatasetWidget.dottedDimensionTable's "issues the ONE read" case asserts a fetch AFTER waiting only on the rendered cell

[yinlianghui]

Observed on an unrelated PR (#4485, the plugin-chatbot DOM-prop whitelist — it touches no dashboard code). Test (shard 1/4) went red on:

FAIL dom packages/plugin-dashboard/src/__tests__/DatasetWidget.dottedDimensionTable.test.tsx
  > renders a LOCAL-only table exactly as the server sent it, and issues the ONE read (#4330)
AssertionError: expected [] to deeply equal [ 'crm_opportunity' ]
- [ "crm_opportunity" ]
+ []
 ❯ packages/plugin-dashboard/src/__tests__/DatasetWidget.dottedDimensionTable.test.tsx:207:23
 Test Files  1 failed | 331 passed (332)

Job: https://github.com/objectstack-ai/objectui/actions/runs/31632075546/job/94233072634 (shard duration 329.91s, import 650.16s — a saturated run).

Not the change under test, and not reproducible in isolation:

• the same file on that branch, run alone: Test Files 1 passed (1) · Tests 6 passed (6);
• re-running the failed job on the identical commit: green (https://github.com/objectstack-ai/objectui/actions/runs/31632075546/job/94235316788).

The race, as the source reads

await waitFor(() => expect(firstRowCells()[0]).toBe('Education'));
expect(requested).toEqual(['crm_opportunity']);

The wait is on the RENDERED CELL, but the assertion is about a metadata READ recorded by installMetaRouter. On the local path the row renders from the server's own rows — it does not depend on that read having resolved or even having been issued — so the two are only incidentally ordered. Under load the render wins and requested is still empty. This is AGENTS.md 测试纪律's profile exactly: an unbounded async step judged inside a window bounded by something else.

Not a timeout bump

AGENTS.md forbids fixing this class by raising a timeout or moving the file into heavyDomTests — both hide the race. The fix is to wait on the thing being asserted, e.g. await waitFor(() => expect(requested).toEqual(['crm_opportunity'])), keeping the rendered-cell assertion as its own step. The "exactly ONE read" property (the point of the case, per #4330) is preserved: a waitFor on set equality still fails on a second read.

Filed unassigned while implementing #4431; not fixed there — out of that card's surface.

[hotlong]

Triage: pm:queue, type: Bug (flaky test). Root cause and fix are both already pinned by the filer: the waitFor is bound to the rendered cell while the assertion is about a separately-resolving fetch — fix is to wait on the thing actually being asserted, not raise a timeout or move the file to heavyDomTests (AGENTS.md forbids both, per the filer). No decision pending.

Size/model suggestion: S, mode:subagent, model: sonnet.

---

Generated by Claude Code

[yinlianghui]

Fresh live occurrence, recorded here rather than as a twin issue — it red-lighted an unrelated PR (#4552, a plugin-grid locale one-liner) on 2026-08-13.

Where: CI run 31680222020, job Test (shard 1/4) attempt 1 — https://github.com/objectstack-ai/objectui/actions/runs/31680222020/job/94383748499

Verbatim:

FAIL  dom  packages/plugin-dashboard/src/__tests__/DatasetWidget.dottedDimensionTable.test.tsx
  > renders a LOCAL-only table exactly as the server sent it, and issues the ONE read (#4330)
AssertionError: expected [] to deeply equal [ 'crm_opportunity' ]
- Expected
+ Received
- [
-   "crm_opportunity",
- ]
+ []
 ❯ packages/plugin-dashboard/src/__tests__/DatasetWidget.dottedDimensionTable.test.tsx:207:23

Shard totals that attempt: 1 failed | 4437 passed | 1 skipped.

Confirmed flake, not a break. Re-running the failed job with zero code change (attempt 2, same head sha 426a4d3a) went green: https://github.com/objectstack-ai/objectui/actions/runs/31680222020/job/94386051468. It also passes in isolation locally (6/6), and main's own Test (coverage) job at 479cc7b4 — that PR's exact base — was green, so it only surfaces under the sharded PR runs.

Mechanism, as this issue already states, confirmed from the source: the case waits only on the rendered cell —

await waitFor(() => expect(firstRowCells()[0]).toBe('Education'));
expect(requested).toEqual(['crm_opportunity']);

— and the file's own amended-pin comment explains why that wait cannot cover the fetch: with no bundle mounted the display equals the authored label, so relabelDimensions returns the server's rows by identity and the cell reads Education on the very first paint. The waitFor is therefore satisfied before the async GET /meta/object/:name has pushed into requested, and the next line asserts an unsynchronized side effect. Under a loaded shard the fetch loses that race; alone on a quiet runner it wins. The sibling case two blocks up is exposed the same way — it asserts ['crm_opportunity', 'crm_account'] after the same single-cell wait, so it is the same defect with a wider target.

Suggested shape (not implemented here — out of scope for #4552): wait on the effect being asserted, e.g. await waitFor(() => expect(requested).toEqual(['crm_opportunity'])), keeping the cell assertion as the separate identity pin it is meant to be.

---

Generated by Claude Code

---

Generated by Claude Code

[yinlianghui]

Second site of the same race, measured today while implementing #4683 (session session_01RnQd8iMMUwXQEV1crFmQiQ). Attaching here rather than filing a duplicate, so one fix closes the class.

packages/plugin-dashboard/src/__tests__/DatasetWidget.dottedDimension.test.tsx:269 — the sibling file of the one this card names — has the identical shape:

await waitFor(() => expect(categoriesOf('bogus')).toEqual(['education']));
// Only the base object was ever asked for.
expect(requested).toEqual(['crm_opportunity']);

in falls through gracefully when the relationship segment is not a relationship. The wait is on the rendered category; the assertion is on installMetaRouter's requested. Same failure text this card quotes:

AssertionError: expected [] to deeply equal [ 'crm_opportunity' ]
- [ "crm_opportunity" ]
+ []
 ❯ packages/plugin-dashboard/src/__tests__/DatasetWidget.dottedDimension.test.tsx:269:23

Same non-reproducibility profile, and measured on BOTH sides of my change so it is not that change:

• full pnpm exec vitest run packages/plugin-dashboard/ with my guard applied: 1 failed | 461 passed (the line above);
• same command with the guard ablated (git checkout origin/main -- AdvancedChartImpl.tsx): 58 passed (58) · 462 passed;
• same command with the guard restored: 58 passed (58) · 462 passed;
• the file alone: green.

So it is load/order dependent exactly as described here, and the remedy this card already prescribes — wait on the thing being asserted, keeping the rendered assertion as its own step — applies verbatim at line 269. Whoever takes this card should sweep the other installMetaRouter call sites for the same pattern rather than fix the two known lines.

---

Generated by Claude Code