Post-merge correctives to #7391: two row-ceiling pins don't grade the cap, two docblocks name the wrong discriminator, changeset level+copy, calendar stale ceiling

[os-project-manager]

Filed from the post-hoc contract audit of merged PR #7391 (verbatim at objectui#7391 comment 5527977938; auditor tier-verified claude-fable-5-1, 75/75 turns, no fallback). That PR landed without its Clause-② review — three consecutive HTTP 529 deaths killed the commissioned review and another actor flipped it ready and merged it — so the review ran afterwards as an audit and its findings are already in main.

This card carries the unforked correctives. The one genuine design fork the audit raised — the five symbols published on @object-ui/react's sole entry, against ruling a′'s "the constant is internal" — is a separate card and is ⛔ not part of this one.

---

⏳ Time-critical: (1) closes at the next release

Both .changeset/ files are still pending on origin/main. Once released, the semver level is frozen into the per-package CHANGELOG and cannot be corrected without a lie or a note. Do (1) first.

(1) Changeset level and copy — .changeset/7210-non-grid-row-ceiling.md

• plugin-gantt / plugin-calendar / plugin-map / plugin-tree declare patch, but they carry the user-visible break: result sets above 2000 rows are no longer fully drawn. Policy is minor-with-the-break-spelled-out. The fixed group makes the released version identical either way — the defect is that the CHANGELOG will file a behaviour break under "Patch Changes". ⇒ flip the four to minor.
• The example reads Showing the first 2,000 of 41,234 records — separators the note never renders. The i18next config carries no format, and the provider-less fallback uses String(v); the unit test's toContain('41234') passing is the proof. ⇒ change the example to the rendered copy: Showing the first 2000 of 41234 records.
• The export list omits the type NonGridCeilingResult. ⇒ add it.

(2) Two row-ceiling pins do not grade the cap they are named for

packages/plugin-map/src/ObjectMap.rowCeiling-7210.test.tsx and packages/plugin-calendar/src/ObjectCalendar.rowCeiling-7210.test.tsx.

Measured: mutating setData(capped.rows) → setData((result as any).data ?? capped.rows) in both — so 2001 rows are drawn while $top and the note stay intact — leaves both pins green, 4/4. They pin the $top and the footnote, not "draws at most N". Ruling a′'s pin is explicit: "the DOM row count equals the ceiling".

⇒ Assert the count handed to the child equals NON_GRID_ROW_CEILING in the above-ceiling case — marker/cluster input for the map, events for the calendar — the way the gantt pin already does via its stubbed child. The mutation above is the ablation: state the prediction first, then show both pins going red under it and green after restore.

(3) Two pin docblocks state a reverse-verification mechanism that is false

• packages/plugin-gantt/src/…/ObjectGantt.rowCeiling-7210.test.tsx:29-33 predicts that deleting $top fails "at the FOOTNOTE assertion". Measured: it fails at the $top assertion (:128). The adapter returns everything, the helper slices, and the note still renders.
• ObjectTree.rowCeiling-7210.test.tsx:24-28 predicts failure "at BOTH the row count and the footnote". Measured: :98 ($top) only — the row count stays 2000.
• The calendar and map docblocks make the same "red at the footnote" claim.

The pins are not vacuous; the stated mechanism is simply wrong, which is worse than silence because the next reader trusts it. ⇒ rewrite to what actually happens: red at the $top assertion; the note still renders because the helper slices whatever arrives.

(4) Calendar external-data sync leaves rowCeiling stale

packages/plugin-calendar/src/ObjectCalendar.tsx:344-347 is the one setData path that does not also setRowCeiling({ truncated: false }). Latent today (ObjectView passes data from mount, so the stale state is never reached), one-line corrective. Pin it: a component that first draws a truncated fetch and then receives a short external data must drop the note.

(5) The false dependency comment — packages/react/src/index.ts:100-104

It reads that @object-ui/react is "the only package all four already depend on". Measured false: packages/plugin-{gantt,calendar,map,tree}/package.json each also list @object-ui/core, @object-ui/components and @object-ui/types.

The real reason was the dispatching seat's same-round barrel fence (#7210 comment 5511301369 — core owned by #7349, types+components by #7182). ⇒ rewrite the comment to state the true reason, and cross-link the decision card below so the next reader finds the open question rather than a settled-sounding rationale.

⚠️ Comment only. ⛔ Do not move, rename or unpublish any of the five symbols in this PR — that is the decision card's scope.

(6) Stale headroom comment

packages/i18n/src/locales/*.ts still says "~1 KB of headroom"; #7399 moved the framework chunk to ~8.3 KB of headroom against its ceiling. One-line refresh.

---

Fences

• ⛔ Nothing in this card touches the published surface of @object-ui/react. Findings ①(b), ② and ③ of the audit — relocating the non-React half to @object-ui/core, retiring NON_GRID_ROW_CEILING_TOP behind a query helper, reshaping NonGridRowCeilingNote to take NonGridCeilingResult — are all on the decision card and are ⛔ out of scope here.
• ⛔ No test may be skipped, disabled or quarantined to get green.
• The audit's finding ⑥ (the merge-reshaped expandFls-7230 assertion) is a resolution, not a weakening — verified in both directions, no FLS regression can pass it silently. ⛔ No action; do not "fix" it.
• Boundary notes recorded by the audit, not defects and not in scope: ObjectMap still runs two queries per load (schema in deps); under ObjectView/ListView the calendar and map draw the host's $top: 100 page and never engage the ceiling; applyNonGridRowCeiling's total is not provenance-safe against a server that omits total (data-objectstack fabricates total = records.length, so such a server would render "first 2000 of 2001").

[os-project-manager]

Claim: session session_01EMrWaQw3XS5DxTHxp4yRyC · branch claude/issue-7507-7391-correctives · seat domain:ui (os-project-manager)

Clause-②: no — nothing here moves contract accept/reject behaviour or the published surface. Path limb: no packages/spec/src/**, no *.zod.ts. Content limb: no symbol added, moved or removed; item (5) is a comment-only edit and the five published symbols are fenced to #7508. The changeset flip patch → minor re-levels an already-decided behaviour change; it does not create one. ⇒ no needs:contract-review, no tier review, lands on CI green.

Provenance. Every item is the auditor's own measurement, not this seat's reading — verbatim at #7391 comment 5527977938, auditor tier-verified claude-fable-5-1 on 75/75 verdict-producing turns with no fallback evidence.

Ordering. Item (1) first and on its own merits: both changesets are still pending on origin/main, and that window closes at the next release. The rest can follow in the same PR.

Related, ⛔ not blocking: #7508 (needs-user-decision) carries the one genuine fork — the five symbols published against ruling a′'s "the constant is internal". This card is deliberately fenced away from it so a shipped-and-red set of cheap correctives is not held hostage to an API decision.

---

Generated by Claude Code

[os-project-manager]

席内核验:通过 —— 并且它订正了本席给的一个数字

domain:ui 执行席,session session_01EMrWaQw3XS5DxTHxp4yRyC。PR #7517,head a221c9288。承重读数自测,不采信报告。

⭐ 派工单里我给的余量数字是错的,实现者分对了

我在单子第 (6) 项写「#7399 把 framework 移到约 8.3 KB 余量」,并要求「自己重测,别抄我的」。它照做了,结果打翻了我:

scripts/check-eager-closure-budget.mjs:474   'i18n-locales': 455_000   ← ceiling
scripts/check-eager-closure-budget.mjs:565   'i18n-locales': 446_076   ← baseline   余量 8,924 B
scripts/check-eager-closure-budget.mjs:475   framework:       71_000   ← ceiling
scripts/check-eager-closure-budget.mjs:566   framework:       61_465   ← baseline   余量 9,535 B

⇒ 我把两个 chunk 压成了一个。#7399 的 A′ 之后 locale catalogue 拿到了自己的 chunk(i18n-locales),而那条陈旧注释说的是 framework —— 这一步我丢了,于是「8.3 KB」既不是 framework 的(9,535 B)也不是 locales 的(8,924 B)。同一个错数字我还写进了 #6939 的一条订正评论,已在那边单独订正(5528562863)。

⚠️ 还有一层精度是我完全没有而它写进注释的:这两个余量是 ceiling 距离它被记录时的 baseline,不是今天构建的实测 —— baseline 之后又有三个 commit 加了 locale key,所以现行余量更小,注释因此指向 pnpm check:eager-closure。要求 dev 重测而不是抄我,这次真的挣到了。

changeset 的 empty-frontmatter 豁免:前提我查了,成立

它用豁免的理由是「所有补正都落在一个尚未发布的变更里」。这条依赖一个可证伪的前提,我查了:

.changeset/7210-non-grid-row-ceiling.md   仍在 origin/main 上   ✓
CONTROL: .changeset/ 共 834 个条目 —— 查询有效

⇒ #7210 的 changeset 确实仍 pending ⇒ calendar 那个 stale rowCeiling 修的是用户从未见过的缺陷,它的 release note 归属就是 #7210 那份 changeset 本身(已在本 PR 内就地订正 patch → minor),不需要第二份。豁免成立。

审计的两条发现,它先复现再修

• map / calendar 的 pin 不判定上限:用审计给的同一处变异(setData((result as any).data ?? capped.rows)),两条 pin 仍然 4/4 全绿,画着 2001 行。复现了才动。
• 修法各自讲清楚了,而且 map 那条不是照抄 calendar:clustering 是 map 自己的防御,会把计数折叠掉(超过 100 个 marker,DOM 里是气泡不是 marker)。enableClustering={false} 让计数可观测而不改变到达 view 的东西(clustering 是 marker 数组的纯函数),于是 DOM marker 计数才字面上是裁定说的「the DOM row count equals the ceiling」。
• 两条 pin 都配了下界用例(calendar 断言 12,map 断言 20),所以一个无条件封顶的实现也满足不了这对断言。这一条我没在单子里要求,是它自己加的,而且正是防 pin 变空转的关键。

⭐ 四个 docblock 的错,它给出了统一解释 —— 这比审计更进一步

审计只说「机制写错了」。它测了四个并给出一个成因:

删掉 $top 之后,adapter 回答的是整个过滤集,所以 applyNonGridRowCeiling 仍然看到多于上限的行,仍然切到上限,仍然从手里的行报出 truncated。画出的行数还是 2000,脚注仍然同时给出两个数字。

⇒ 所以丢 $top 是带宽回归,不是正确性回归 —— 而这恰好解释了为什么 $top 断言不与脚注断言冗余。审计说的是「docblock 写错了」,它答的是「为什么这条断言必须单独存在」。后者才是下一个读者需要的。

Clause-② no:复核确认

「零导出符号增删改名」是它的声明,与本 PR 的围栏一致(审计的 ①(b)/②/③ 全留给 #7508)。第 (5) 项只改注释文字,导出语句未动。⇒ no 成立,⛔ 不挂 needs:contract-review。

一条诚实的 NOT MEASURED,记一笔

check:readme-exports 在这棵树上 exit 1,verdict 是它自己的 population COLLAPSED + dist/index.d.ts is not on disk。它报成 NOT MEASURED 而不是 green,也不是 failure —— 那是前置条件不满足,不是本 diff 的发现,而且该 gate 判的是 README 与导出面的对应,本分支不动任何导出,两个方向都红不了。#7510 的实现者对同一个 gate 做了同样的诚实处理,而 CI(先构建)在它那里判了 success。 这条会在本 PR 的 CI 上同样得到真读数。

处置

CI 于 16:08 起跑,尚无结论。全绿即剥 draft 放行。Fixes #7507。

⚠️ mergeable_state 现读 behind —— base 是 0e3b3be09(即当前 main),而合并队列里还排着 #7473 → #7511 → #7514。等队列排空后复看,⛔ 现在不做任何 rebase/merge 动作。

---

Generated by Claude Code

[claude]

os-dev-report

{
  "issue": 7507,
  "status": "done",
  "branch": "claude/issue-7507-7391-correctives",
  "pr": "https://github.com/objectstack-ai/objectui/pull/7517",
  "premise_still_valid": true,
  "summary": "All six correctives landed; the seventh strand (the five published symbols) was not touched — this branch adds, moves, renames and retires ZERO exported symbols, so the Clause-2 determination stands. Nothing was taken on the audit's say-so: every claim was re-derived first, and the audit reproduced in full. (1) CHANGESET — re-measured the rendered footnote with a throwaway probe that rendered NonGridRowCeilingNote three ways and asserted a deliberately WRONG string as its positive control (3 failed, so the render really ran): provider-less and I18nProvider paths both print 'Showing the first 2000 of 41234 records. Narrow the filter.', unknown-total prints 'Showing the first 2000 records. Narrow the filter.' — no separators, because i18n.ts declares no `format` and fallbackInterpolation.ts interpolates via String(v). Probe deleted. Changed: plugin-gantt/calendar/map/tree flipped patch->minor with a sentence saying why, the example replaced with the rendered copy, NonGridCeilingResult added to the export list (it is already exported at packages/react/src/index.ts, so this is documentation not a new export). (2) MAP + CALENDAR PINS — first reproduced the audit exactly: with setData(capped.rows) -> setData((result as any).data ?? capped.rows) in both components, both pins stayed GREEN 4/4 while 2001 rows were drawn. Added the missing assertion: calendar stubs CalendarView via importOriginal (the gantt/unscheduled-7071 idiom) and asserts data-event-count === NON_GRID_ROW_CEILING above the ceiling and 12 below; map gets a case of its own with enableClustering={false} — clustering folds the count into bubbles, and it is a pure function of the marker array, so turning it off changes what is on screen and not what reached the view — asserting the DOM map-marker count equals the ceiling, plus 20 markers in the below-ceiling case with no prop. PREDICTED RED SET (stated before the run): calendar rowCeiling red at the new count assertion, map rowCeiling red at the new marker count, externalDataCeiling-7507 red at its count waitFor. MEASURED: exactly that — 3 failed / 4 passed, at ObjectCalendar.rowCeiling-7210.test.tsx:124, ObjectMap.rowCeiling-7210.test.tsx:125, ObjectCalendar.externalDataCeiling-7507.test.tsx:120, each 2001 against 2000. (3) DOCBLOCKS — re-measured all four by deleting $top: NON_GRID_ROW_CEILING_TOP from each component. PREDICTED: all four red at the $top assertion only, because the adapter returns everything and applyNonGridRowCeiling still slices from the rows in hand. MEASURED: gantt :128, tree :98, calendar :98 (pre-fix), map :94 — all at `expect(params.$top).toBe(NON_GRID_ROW_CEILING_TOP)`, never at the footnote, and the tree's row count stayed 2000. Re-measured calendar and map AFTER their count assertions landed, as asked: unchanged, still $top-only, now at :115 and :94 — so those two docblocks did NOT become true, and all four were rewritten to the measured mechanism plus, for calendar and map, the second ablation their new count assertions exist for. (4) CALENDAR STALE rowCeiling — confirmed ObjectCalendar.tsx's external-data sync was the one setData path with no setRowCeiling reset (the other four, and both of ObjectMap's, carry it). Fixed with that one line and pinned in the new ObjectCalendar.externalDataCeiling-7507.test.tsx, whose first assertion is a live control proving the note really was raised by this component's own fetch. PREDICTED RED SET: the new pin red at expect(screen.queryByRole('note')).toBeNull(), rowCeiling-7210 green. MEASURED: exactly that — 1 failed / 3 passed, at ObjectCalendar.externalDataCeiling-7507.test.tsx:131. (5) FALSE DEPENDENCY COMMENT — verified false from the manifests: plugin-gantt/calendar/map/tree each also list @object-ui/core, @object-ui/components and @object-ui/types (i18n is NOT common to all four — plugin-map lacks it). Comment rewritten to the real reason (the round's barrel fence over exactly those three, plus @object-ui/react having no subpath export) and cross-links #7508 as the open decision. Comment text only; the export statements are byte-untouched. (6) HEADROOM COMMENT — re-measured live from the gate's own exported constants rather than copying the dispatch's figure: i18n-locales ceiling 455,000 over baseline 446,076 = 8,924 B (framework is 71,000 over 61,465 = 9,535 B, not the 8.3 KB the dispatch quoted). Since #7399 the locale bytes are budgeted by `i18n-locales`, not `framework`, so the old comment named the wrong line as well as the wrong size. All ten packs rewritten to name the chunk, state the ceiling's distance from the baseline it was measured from, and point at `pnpm check:eager-closure` for the figure in force. DECLARED LIMITATION: this is not a fresh console build. Three commits have added locale keys since that baseline was recorded, so today's live headroom is somewhat smaller; the comment is worded as the ceiling-to-baseline distance, which is exactly true, rather than as today's measurement. A fresh figure needs a full apps/console build — a repo-level scan CI's Bundle Analysis owns on this PR. NOTHING FAILED TO REPRODUCE: every measurement in the audit that this card acts on came back identical. One dispatch figure did not match the tree (the 'roughly 8.3 KB' framework headroom; measured 9,535 B), and per instruction I wrote what I measured. CHANGESET DISCIPLINE: .changeset/7507-non-grid-ceiling-correctives.md uses the empty-frontmatter exemption deliberately — every corrective here lands inside a change that has NOT been released (the #7210 changeset is still pending on main), so the calendar fix repairs a defect no user has seen, and the release note it belongs to is the #7210 changeset, corrected in place. check-changeset-presence.mjs accepts it explicitly.",
  "tests": "Ablation preconditions: vitest's root config aliases every @object-ui/* specifier to that package's src/, so no dist sits in the resolution path and these ablations need no rebuild — stated because the usual rebuild-and-preflight discipline is conditioned on dist resolution, which does not hold here. Every mutation was proven on disk by grepping BOTH the injected marker (count 0->1) and the removed text (count 1->0) before the run; every restore was proven byte-identical by comparing `git hash-object` against the path's HEAD blob (empty read treated as failure), and each script carried `trap restore EXIT INT TERM` with absolute paths. All heavy runs went through `bash scripts/pm/os-verify-lock.sh -c ...` with OS_VERIFY_LOCK_SLOT=dev-7507 set before the first attempt; verdicts read from the printed VERDICT line, never a bare $?; every capture redirected to a file BEFORE the exit code was read (no pipes). BASELINE at origin/main a27d153c2: 5 files / 16 tests passed. AUDIT REPRO (pre-fix, uncapped hand-off on calendar+map): 'Test Files 2 passed (2) | Tests 4 passed (4)' — the audit's finding that both pins stay green at 4/4 while drawing 2001 rows, reproduced exactly. $TOP ABLATIONS (pre-fix): gantt 1 failed/2 passed at ObjectGantt.rowCeiling-7210.test.tsx:128 'expected undefined to be 2001'; tree 1 failed/1 passed at :98; calendar 1 failed/1 passed at :98; map 1 failed/1 passed at :94 — all four at the $top assertion, none at the footnote. All four restored byte-identically (RESTORE-OK with the HEAD blob hash printed). POST-FIX ABLATION A (uncapped hand-off): 'Test Files 3 failed (3) | Tests 3 failed | 4 passed (7)' at ObjectCalendar.rowCeiling-7210.test.tsx:124, ObjectMap.rowCeiling-7210.test.tsx:125, ObjectCalendar.externalDataCeiling-7507.test.tsx:120 — matches the prediction stated before the run. POST-FIX ABLATION B (drop the external-data rowCeiling reset): 'Tests 1 failed | 3 passed (4)' at ObjectCalendar.externalDataCeiling-7507.test.tsx:131 — matches the prediction. POST-FIX ABLATION C ($top deleted, for the docblocks): 'Tests 2 failed | 3 passed (5)' at ObjectCalendar.rowCeiling-7210.test.tsx:115 and ObjectMap.rowCeiling-7210.test.tsx:94, both 'expected undefined to be 2001' — still $top-only, matching the prediction. FINAL UNION at HEAD a221c9288 (git rev-parse --short HEAD from that run): `pnpm exec vitest run packages/plugin-calendar packages/plugin-map packages/plugin-gantt packages/plugin-tree packages/react/src/utils/nonGridRowCeiling.test.tsx packages/i18n --maxWorkers=2` => 'Test Files 170 passed (170) | Tests 1687 passed (1687)', VERDICT command-exit 0. TYPE-CHECK at a221c9288: `pnpm exec turbo run type-check` over the six touched packages with --concurrency=2 (deps built first) => 'Tasks: 21 successful, 21 total'; all six type-check tasks verified as 'cache miss, executing' rather than zero-matched, and each package's script is `tsc --noEmit && tsc -p tsconfig.test.json` so the test files are covered — confirmed with `tsc -p tsconfig.test.json --listFiles`, which lists the new pin (1 hit, not 0). GATES at a221c9288, each redirected then read: check:control-bytes 0 ('scanned 6196 tracked text file(s)'), check:vi-mock-specifiers 0, check:vi-mock-inherit 0, check:i18n-keys 0, check:i18n-drift 0, check:i18n-dead-keys 0, node scripts/check-changeset-presence.mjs 0 ('17 source file(s) of 6 released package(s) changed, and this change declares 1 changeset(s)'). Own control-character sweep over the changed set: clean. LINT — DECLARED NARROWING, with the three pieces of evidence. (a) Population from eslint's own config, not a guess: eslint.config.js's one files entry is '**/*.{ts,tsx}', so 17 of the 19 changed files are in scope (the two .changeset/*.md are not). (b) File count from --format json: 17 results, i.e. every in-scope changed file was judged; 0 errors, 103 warnings, all @typescript-eslint/no-explicit-any in the fixtures' `as any` idiom, and lint.yml sets no --max-warnings so that rule cannot fail CI. (c) Invariance for untouched files: the config declares no parserOptions.project / projectService, so type-aware linting is NOT enabled — a file's verdict is a function of that file's own source and the shared config, and this diff moves neither for any untouched file. `pnpm lint` runs the whole farm on this PR regardless. NOT MEASURED (not green, not red): check:readme-exports exits 1 on this tree with 'its type entry ./dist/index.d.ts is not on disk -- run pnpm build first' and a 'population COLLAPSED' verdict — a prerequisite failure needing every package built, which is CI's run. It cannot go red on this diff either way: it judges README docs against the export surface, and this branch changes no export. check:eager-closure likewise needs a full apps/console build and is Bundle Analysis's job on this PR. CHANNEL NOTE: the container's repo-scoped REST is open on this session (branch push, PR create 201, PR body read-back), but /search/* is refused ('sessions are bound to their configured repositories'), so the out-of-scope duplicate check used one targeted MCP search_issues — it returned 10 hits, so the non-empty result validated itself and no control query was needed.",
  "mcp_calls": "2 — one search_issues (duplicate check for the out-of-scope finding) and one issue_write (filing #7518). Everything else, including the branch push, the draft PR, its body read-back and this report comment, went through the container's repo-scoped REST or plain git.",
  "open_questions": [],
  "out_of_scope_findings": [
    "filed as #7518: scripts/check-eager-closure-budget.mjs states two different aggregate headrooms 200 lines apart — the objectui#7399 block's '11,988 bytes' contradicts the same file's 45,686 (= MAX_EAGER_CLOSURE_GZIP_BYTES 3,268,000 - BASELINE.gzipBytes 3,222,314). Not fixed in passing: the right correction is a judgement (restate the figure, the ratio, or defer to the live verdict), and the file is in performance-budget.yml's runtime closure, so touching it costs a console build and owes its own test suite. Unassigned, unlabelled, for PM triage."
  ]
}

---

Generated by Claude Code