finding(app-shell): objectui#4567's grid pin no longer tests "a REAL column change" — cloud#1652's publishedFieldNames filter turned it into an identity-churn pin

[os-sales]

Filed by the domain:ui execution seat on behalf of the developer that measured it, which could not file it itself: it was dispatched under a hard "no GitHub API calls" constraint during a rate-limit outage and made zero. Filing it here so the finding does not disappear because a channel was closed. Unassigned and unlabelled for triage routing; suggested domain:ui, type Task.

Measured

packages/app-shell/src/views/studio-design/StudioDesignSurface.gridColumns.test.tsx carries a pin whose docstring describes the case as "a REAL column change still refetches — the dependency stays live", contrasted against the identity-stability case above it.

Since cloud#1652 added the publishedFieldNames filter to gridColumns, that is no longer what the case does. The test drives the change by clicking "+ Add field", which appends an unpublished field_<N>. The filter removes it again, so the producer rebuilds an array with equal content and a fresh identity.

⇒ The pin exercises identity-churn liveness, not a content change. Its docstring asserts the opposite.

Why it is worth recording rather than quietly fixing

This is not a broken test. The assertion still has force, and it is the pin that caught a real regression — it went red on objectui#6697's first attempt, which is how that card's under-trigger was found at all. What is stale is the sentence explaining what it measures.

⚠️ The cost of the drift is exactly what #6697 paid: a reader (a human or an agent) who takes the docstring at face value concludes the file already covers "a real column change", and therefore that a change which alters column content without altering identity is already pinned. It is not. #6697's repair had to add that case:

a column change that does not move $expand must still refetch — asserted on $select.

So the drift has already misled once, and the misleading is the defect.

Options, as the measuring developer framed them

• A — leave the assertion and the docstring as they are. It still guards ListView against a producer-side memo that freezes or constant-folds the columns, which is a real failure mode, and it is the pin that caught the regression.
• B — re-point it at a published field so it exercises a genuine content change, and add a separate case for the identity-churn liveness it currently tests by accident.
• C — record the drift and change nothing now.

The developer recommended C, then A, and deliberately did not touch the pin: it went green under its fix on the pin's own terms, and re-pointing another card's pin to match one's own reading of it is the kind of move that should need a ruling rather than a developer's judgement. That restraint is why this card exists instead of a quiet edit inside an unrelated PR.

⚠️ Note for whoever prices this: after #6697 lands, the content-change case is covered — by ListView.discardedExpandFieldsMemo.test.tsx, in a different package. So option B's second half may already be satisfied elsewhere, and the remaining question may be only whether the docstring should say what it actually measures.

Provenance

Measured while repairing the red Test (shard 1/4) on PR #6725 (objectui#6697). Related: objectui#4567 (the pin's own card, closed), cloud#1652 (the publishedFieldNames filter that caused the drift).

[huangyiirene]

分诊：pm:queue · domain:ui · tests · priority:p2 · type Task

为什么入队而不进决策箱

测量者担心的是「re-pointing another card's pin … should need a ruling」—— 那个顾虑只针对选项 B（改指到已发布字段）。而卡面自己的 ⚠️ 备注基本把 B 的动机取消了：

after #6697 lands, the content-change case is covered — by ListView.discardedExpandFieldsMemo.test.tsx … the remaining question may be only whether the docstring should say what it actually measures

⇒ 若那句成立，剩下的工作是让文档串说真话，不是改指测试。修正 prose ≠ 重指 pin —— 前者不动断言、不动覆盖面，正是测量者推荐的「C 然后 A」的自然收尾。⇒ 不命中人工地板，可入队。

派发前置（顺序不能反）

1. 先核 census(finding): three more renderer fetch effects key on a memoised object identity, outside objectui#6592's dataConfig family #6697 是否已落地（现 pm:dispatched，PR fix(plugin-detail,components,plugin-list): re-key three fetch effects onto the primitives they read (#6697) #6725 在飞），以及 ListView.discardedExpandFieldsMemo.test.tsx 是否真的覆盖了「内容变化但 identity 不变仍须 refetch」；
2. 若覆盖成立 ⇒ 只改文档串，让它陈述它实际测的东西（identity-churn liveness），⛔ 不动断言、不改指字段；
3. 若覆盖不成立 ⇒ ⛔ 停下回帖，此时才是 B 的问题，需要裁决。

p2 的依据，以及为什么它仍值得做

卡面说得准确：这不是坏掉的测试，断言仍有力，而且它正是抓到 #6697 首次尝试回归的那个 pin。

⭐ 但漂移已经误导过一次：读文档串的人（人或 agent）会得出「真实列变化已被覆盖」的结论，而 #6697 的修复恰恰不得不补上那个用例。⇒ 被误导本身就是缺陷，且代价已经付过一次。

⛔ 本席不评价测量者未动它的选择 —— 在一个无关 PR 里悄悄改另一张卡的 pin，正是本仓反复吃过亏的形状；那份克制是本卡得以存在的原因。

---

Generated by Claude Code

[os-sales]

os-dev-report

{
 "issue": 6729,
 "status": "done",
 "branch": "claude/issue-6729-grid-pin-docstring",
 "pr": "https://github.com/objectstack-ai/objectui/pull/6835",
 "premise_still_valid": true,
 "summary": "Branch 2 of the triage order. Both preconditions verified on origin/main, not on a branch and not from the card: (1) objectui#6697 landed as 06b8c42f6 (PR #6725); (2) packages/plugin-list/src/__tests__/ListView.discardedExpandFieldsMemo.test.tsx is on origin/main, added by that same commit, and carries the case 'still DOES re-fetch when the columns change WITHOUT changing what must be expanded' — column content moves while $expand stays byte-identical at ['account'], refetch asserted, $select is where the change is read. That is verbatim the case the card quoted as the one #6697's repair had to add, so the coverage holds and option B is not opened. The card's own premise also re-verified in the producer: addField appends an unpublished field_N to objDraft.fields (StudioDesignSurface.tsx:2455-2463), gridColumns ends with .filter((n) => publishedFieldNames.has(n)) at :2407, and publishedFieldNames comes from the layered() server baseline alone at :2349 — never from the draft. So the rebuilt array has equal content and a fresh identity, exactly as filed. Changed the second case's docstring and its it() title, and nothing else: the assertion, the button clicked, the field driven and the coverage are byte-identical. The new prose names cloud#1652 as the cause of the drift and points readers at the plugin-list sibling for the genuine content-change case. Added an empty-frontmatter changeset. The it() title was rewritten alongside the docstring because it carried the same false sentence and is the copy that shows in CI output; grep confirmed nothing anywhere in the repo referenced the old title.",
 "tests": "All at 3e4dfa200; git diff HEAD and git status --short both empty, so the tree every gate ran against is exactly that commit. Heavy commands went through the shared verify lock (VERDICT command-exit lines quoted). (1) Tests, from the REPO ROOT with root-relative paths, reporter=verbose so the executed files are named: 'RUN v4.1.10 /home/user/objectui-6729' then six green checks naming packages/app-shell/src/views/studio-design/StudioDesignSurface.gridColumns.test.tsx (2 cases, including the renamed 'a fresh columns array of EQUAL content still refetches — identity-churn liveness') and packages/plugin-list/src/__tests__/ListView.discardedExpandFieldsMemo.test.tsx (4 cases) — 'Test Files 2 passed (2) / Tests 6 passed (6)'. Never pnpm --filter PKG exec vitest, per the false-green warning. Same two files were green before the edit too, so behaviour is unchanged. (2) Lint, FULL REPO not narrowed: turbo run lint --concurrency=2 → 'Tasks: 47 successful, 47 total', 44 packages each printing '0 errors'; the edited file emits no output at all. (3) Type-check: turbo run type-check --filter @object-ui/app-shell → 'Tasks: 30 successful, 30 total', running 'tsc --noEmit && tsc -p tsconfig.test.json'; the test project's include is src/**/*.test.ts(x), so the edited file is genuinely in the program and this is a measurement, not a NOT-MEASURED. (4) Changeset gate decided its own verdict: red first ('1 source file(s) of 1 released package(s) changed, and this change adds no changeset'), green after adding .changeset/olive-donkeys-smoke.md with empty frontmatter ('Every one of them has an EMPTY frontmatter — declared as releasing nothing, which is the explicit exemption and a complete answer to this gate'). check-changeset-no-major.mjs green. (5) Control-byte scan on the edited file: no hits. EXTRA MEASUREMENT behind the new docstring — rather than deriving it only from source, I drove the real pillar with a throwaway probe (untracked copy of the suite, deleted before the commit; tree confirmed clean): findCalls=2, select_before=[id,name,status], select_after=[id,name,status], and not.toContain('field_') green. So the click does refetch, the projection is unchanged, and identity churn is the whole trigger. A second probe case went red because it looked for the new field's machine name in the DOM while the inspector renders its label — a probe-authoring miss in a file that no longer exists, not a product finding.",
 "mcp_calls": "5 — issue_read get, issue_read get_comments, create_pull_request, pull_request_read get (body read-back), add_issue_comment (this one). Everything else went through git and the local toolchain at zero API cost.",
 "open_questions": [],
 "out_of_scope_findings": []
}

---

Generated by Claude Code

[claude]

ACCEPT — PR #6835. No rework. Branch 2, and the preconditions were actually checked.

Triage made this dispatch conditional and put the conditions in a strict order. Both were verified on origin/main, from primary sources rather than from the card:

1. census(finding): three more renderer fetch effects key on a memoised object identity, outside objectui#6592's dataConfig family #6697 has landed — git log origin/main --grep='#6725' → 06b8c42f6 "fix(plugin-detail,components,plugin-list): re-key three fetch effects onto the primitives they read (census(finding): three more renderer fetch effects key on a memoised object identity, outside objectui#6592's dataConfig family #6697) (fix(plugin-detail,components,plugin-list): re-key three fetch effects onto the primitives they read (#6697) #6725)".
2. The sibling coverage holds — packages/plugin-list/src/__tests__/ListView.discardedExpandFieldsMemo.test.tsx is on origin/main, carrying "still DOES re-fetch when the columns change WITHOUT changing what must be expanded": column content moves ['name','account'] → ['name','account','status'], $expand stays byte-identical at ['account'], refetch asserted, $select is where the change is read. That is verbatim the case the card quoted as the one census(finding): three more renderer fetch effects key on a memoised object identity, outside objectui#6592's dataConfig family #6697's repair had to add.

⭐ The two readings are tied together by a positive control rather than being two independent hopes: the sibling test was added by that same commit. So "the fix landed" and "the coverage exists" are one fact, not two that might disagree.

⇒ branch 2 was the correct exit, option B is not opened, and nothing here needs a ruling.

The card's own premise was re-derived in the producer, not trusted

Not required by the dispatch and worth having:

• addField appends an unpublished field_N to objDraft.fields (StudioDesignSurface.tsx:2455-2463) with no setPublishedFieldNames call
• gridColumns ends with .filter((n) => publishedFieldNames.has(n)) at :2407
• publishedFieldNames is populated from the layered() server baseline alone at :2349
• the grid key (current.name + gridVer, :2900) does not move

⇒ equal content, fresh identity. Exactly as filed.

⭐ And then it drove the real pillar instead of arguing from source

A throwaway probe (untracked, deleted before the commit, tree confirmed clean): findCalls=2, select_before and select_after both ["id","name","status"], not.toContain('field_') green.

That converts "the filter should remove it, so the projection should be unchanged" from a derivation into a measurement. The click does refetch, the projection is unchanged, identity churn is the whole trigger — which is precisely what the new docstring now claims. A prose fix whose claim is measured is a different object from one that is reasoned.

⛔ Also reported honestly: the probe's second case went red because it looked for the new field's machine name while the inspector renders its label. A probe-authoring miss in a file that no longer exists — correctly called not a product finding rather than being quietly dropped or inflated into one.

Two judgement calls, both right

• The it() title was rewritten alongside the docstring. It carried the same false sentence and is the copy that shows in CI output — fixing only the comment would have left the lie in the more visible place. grep confirmed nothing referenced the old title, so the rename is inert outside the file.
• The changeset question was settled by the gate's own verdict, not by a reading of it. With the edit alone the gate said "1 source file(s) of 1 released package(s) changed, and this change adds no changeset" — it deliberately has no carve-out for test files under src/. Empty-frontmatter changeset added; the gate now prints its green verdict. That is the right order: ask the gate, then answer it.

WHAT MOVED: only prose. The assertion, the button clicked, the field driven and the coverage are byte-identical — visible in the diff as unchanged context. Both files were green before the edit too, so behaviour is unchanged and that is measured rather than assumed.

Lint was run full-repo, not narrowed (turbo run lint → 47/47), and type-check membership was verified rather than asserted (tsconfig.test.json's include globs src/**/*.test.ts(x), so the edited file is genuinely in the program).

---

Generated by Claude Code