refactor(plugin-grid): retire the redundant ObjectGridColumnHolds.headerIcon hold - #6686

Part of #6424

⚠️ Part of, not the closing keyword, and that is a deliberate deviation from the dispatch order. The order asked for the closing keyword on this card, on the premise that all three items would land. (That keyword is deliberately not spelled next to the card number anywhere in this body: GitHub's parser matches it regardless of any negation around it, and this card must stay open.) Item 1 did not land — its precondition is measurably absent on this ref (evidence below). Closing the card would bury an unfinished, already-ruled item, and the inbox filter only reads open cards. The card stays open for item 1.

Two of the three ruled items land here, both measured rather than inherited.

Item 1 — the emit-side fitContent cast: STOPPED, with evidence

The line, located on my own ref

Both numbers in circulation are stale. Measured at base 9d86e1d79:

Three independent measurements say do not remove it yet.

1a. The ruling's precondition is not met — PR #6673 has not merged

The ruling ties this cast to the fitContent declaration: it "travels with this verdict". That declaration is still in flight — PR #6673 is state=open, merged=false.

Measured on my ref, each zero with a positive control in the same query shape:

1b. The cast is a no-op, but NOT for the reason the ruling assumed

Read via the TypeScript checker at the site, cast stripped:

line 3606  expr=`(col as any).fitContent`
   type of `col` (cast stripped) = any   isAny=true

applyColumnChrome = (col: any) => makes orderedColumns an any[], and the loop adds a second as any[]. So the cast never compensated for the undeclared key — it compensates for nothing. This is exactly the lesson PR #6673 measured on data-table.tsx ("cast count is not the instrument, the normalization is"), reproduced on the emit side. Removing it buys zero type safety, now or after #6673 lands.

1c. It is load-bearing as another guard's only real-file control

columnReadBoundary-6458.test.ts uses this exact cast as its Anti-vacuity control 3, and its own comment says: "If that read is ever retired too, this control must be re-pointed at another real cast — never deleted, and never re-pointed back inside the region."

Measured with the test's own scanner regex over the whole file: (col as any).fitContent at :3606 is the only (col as ...) cast read in ObjectGrid.tsx. There is no other candidate to re-point at.

Observed, not predicted — ablation with the mutation proven on disk and restoration observed:

anchored BEFORE: '(col as any).fitContent'=1     blob 0905b2cc
anchored AFTER:  '(col as any).fitContent'=0     blob e19e6a6e   MUTATION PROVEN ON DISK
 FAIL  columnReadBoundary-6458.test.ts > the scanner finds a cast read in the REAL file, outside the guarded region
 AssertionError: expected [] to include 'fitContent'
 Test Files  1 failed (1) | Tests  1 failed | 7 passed (8)
restore: git diff HEAD empty, blob back to 0905b2cc

Baseline on the clean tree for comparison: Test Files 1 passed (1), Tests 8 passed (8), EXIT=0.

⇒ Removing the cast today costs a live guard its only real-file anti-vacuity control and buys nothing. This needs the seat, not me — it is a re-point decision plus an ordering question against #6673.

Item 2 — ObjectGridColumnHolds.headerIcon removed. MEASURED, not inherited

The order was explicit that the previous round's type-algebra derivation must not be carried forward. It was not. The probe confirms the derivation — the ruling stands.

The ruled probe

Run through the same program the test project uses (tsconfig.test.json, paths: {}, so @object-ui/types resolves through the built .d.ts):

The two controls are the point: a probe that can only answer true measures nothing.

The emit-type ablation — the claim that had never been observed

The actual claim is "the emit types are unchanged with the member absent". Instrument: the TypeScript checker resolves both emit types and prints every member with its resolved type, sorted, to JSON — a structural snapshot, not a derivation. Each leg proves its mutation on disk (anchored grep counts plus a moved blob hash) and its restoration by observation (git diff HEAD empty, hash back), under a trap ... EXIT INT TERM with absolute paths, restoring via an explicit git checkout HEAD -- absolute-path.

The positive control is what makes the identical result a measurement: the instrument demonstrably detects a holds-member removal that matters. The landed change re-snapshots to IDENTICAL against the pre-edit baseline.

The three facts, re-confirmed with controls rather than quoted

Note on the read count: the card body says "2 reads", the ACCEPT says "1". Both are right at different granularities — 1 render site, 2 syntactic col.headerIcon occurrences (a guard and the render). Stated precisely here so the next reader does not have to re-derive it.

The pinned contrast, stated explicitly — and untouched

pinned stays load-bearing on both counts, which is exactly what makes headerIcon's removal safe and pinned's unsafe:

Two different routes to the same "not in the band", and they must not be conflated — that conflation is how a measurement blurs the two keys. pinned is untouched by this PR.

⚠️ The removal MOVES a liveness claim rather than deleting one

While the hold existed, the emit types declared headerIcon on their own account. They now get it from TableColumn alone, so reverting #6615 would drop it from both emit types in silence while ObjectGrid goes on writing it at three sites. That is a new exposure created by this change, and the new pin's first assertion covers it — it is the guard the deleted member used to provide implicitly.

The pin is not vacuous. Simulating that revert turns it red, mutation proven on disk and restoration observed:

src/__tests__/columnHoldsExpiry-6424.test.ts(65,29): error TS2344: Type 'false' does not satisfy the constraint 'true'.
tsc EXIT=2      restore: git diff HEAD empty, blob back to eebaadd3

And the pin really is compiled — tsc -p tsconfig.test.json --listFiles: columnHoldsExpiry-6424.test.ts → 1 (control: columnEmitBoundary-6004.test.ts → 1; negative control zzNoSuchFile → 0). Green from a file the compiler never read would be worth nothing.

Item 3 — the docblock's general claim, corrected

The claim that the held keys are "undeclared by TableColumn" is false as written since #6615. PR #6651 corrected only the one sentence it was already editing; this PR owns the rest:

• ObjectGrid.tsx verdict block — headerIcon moves from HELD to DECLARED, with the expiry and the measurement that retired the hold recorded.
• ObjectGridColumnHolds docblock — names "undeclared by TableColumn" as the interface's entry condition, flags that it is a claim about another package that can lapse with nothing going red, and states the rule: re-check per key when the owning card closes, never inherit. Plus the ⛔ that a key whose only declaration is this interface is not in that position.
• the pinned member doc — now records both counts that make it load-bearing, and the 27→26 measurement.
• columnEmitBoundary-6004.test.ts — its docblock still said the interface declares two keys and its test title said "the two held keys". Both false after this change; corrected, and the routes are now pinned in the new file rather than described in prose.

Gate verdicts — each exit code captured before any pipe, each quoting the gate's own line

Union re-run after the final commit, at 83ad9ffec:

Declared narrowing on lint

Repo-wide pnpm lint is turbo run lint across 40 packages; I ran the one package this diff touches. Three pieces of evidence, not two:

1. Population read from eslint's own config, not my guess: eslint itself selected 134 files for packages/plugin-grid.
2. Count from --format json: 134 files; my three changed files at 0 errors (ObjectGrid.tsx 214 warnings, all pre-existing; the two test files 1 and 0).
3. Configuration invariance: eslint.config.js has no projectService / parserOptions / project: — grep exit 1, no matches — so type-aware linting is off and this diff cannot move the verdict of any file it does not touch.

⚠️ One honest note: adding --no-inline-config (an objectstack recipe, not this repo's) surfaces 3 errors — in demo/main.tsx, predicate-surface-parity.test.tsx and BulkActionDialog.tsx, none of them in this diff. This repo's lint script is eslint . without that flag, which is what CI runs and what returned 0.

Changeset

Empty frontmatter — the gate's own explicit exemption, quoted from its help text: "If this change really should release nothing, say so — that is a pass, not a workaround." Justified by measurement, not assertion: ObjectGridColumnHolds is not part of the package entry's exported surface — 0 occurrences in the built packages/plugin-grid/dist/index.d.ts (positive control: ObjectGridColumnState, which the entry does re-export, → 1), and headerIcon → 0 there too. Runtime behaviour is untouched; this is types, comments and tests.

Scope

packages/types, packages/components/src/renderers/complex/data-table.tsx, app-shell, i18n, sdui-parser, plugin-list and @objectstack/spec are all untouched — git diff --name-only is exactly the three packages/plugin-grid files plus the changeset.

Generated by Claude Code

Generated by Claude Code