fix: one engine emission, structural brands, and a verified wrap-rows exit

[blove]

Summary

The last three follow-ups in our lane. Two turned out to be bigger than filed.

---

1. A branded type was nominally split — and five brands were affected, not one

pretable-model.ts carried an as unknown as because PretableGroupId's unique symbol was emitted once by tsc into grid-core/dist and again by tsup into core/dist — two unique symbols from two declaration files are two different types.

The audit found five brands crossing that boundary, and one of them was worse than a cast:

brand	status
groupIdBrand	live — cast #1
rowModelDescriptor	live — cast #2, plus a silent failure
columnDescriptor	latent (required key would have blocked)
gridUiCoreType	latent-weak (optional key — variance marker meaningless across the seam)
rowLayoutControllerType	not published today; converted so no bad exemplar remains

RowOf<> / RowIdOf<> / ColumnsOf<> match structurally on rowModelDescriptor, so across the seam they resolved to never with no diagnostic at all. Strictly worse than a cast, which at least marks the spot.

Approach: string-literal brand keys (readonly "~pretableGroupId"). Structural, so N emissions are one type regardless of future bundling topology. The "re-export the declaration" alternatives were rejected because @pretable-internal/* are private and unpublished — both would require publishing an internal engine's entire surface.

Brands were only half of it

TypeScript relates a deferred conditional type by the identity of its alias declaration, and PretableAggregateOutputOf<> is one — no brand change can reach that. So the engine now has a single emission: @pretable/react reaches it through ɵcreateGridUiCore re-exported from @pretable/core, rather than importing @pretable-internal/grid-core directly.

Independently verified: react/dist/index.mjs defines createGridUiCore zero times and imports it from '@pretable/core', where it is defined once. Two runtime copies became one.

A correction to the brief

I said typecheck:public would expose the split. It would not — react/dist/index.d.ts imports PretableGroupId from @pretable/core rather than re-emitting it, so consumer-level type tests saw one consistent brand. The split lived in react's own dist-resolved compile and in the .api.md reports, which printed readonly [groupIdBrand] while never declaring groupIdBrand — not even self-consistent. Assertions went where the defect is, plus a fail-closed report guard.

pretable-model.ts casts: 8 → 5; as unknown as: 2 → 0. The five that remain are unrelated local narrowings.

2. A hand-copied union that had already broken the build

moveFocus's movement parameter was a hand-written copy of the string union. When first-column/last-column were added, the stale copy rejected them and broke the build. It now imports PretableIndexedFocusMovement; the cast at the call site is gone.

Pinned so the drift cannot return: adding a member to the engine union leaves pretable-model.ts with an empty diff and fails only the enumeration assertion.

3. The wrap-rows claim was true — the summary above it was not

wrap-rows was a keyboard trap until recently (WCAG 2.1.2, 120 presses without escape), so "Tab leaves in one press" is the claim that the trap is gone — and it was the one configuration nothing on the site rendered.

A fixture now drives it. Chromium and WebKit produced byte-identical traces:

from	Tab forward	Shift+Tab back
top-left cell (entry)	12	1
mid-grid	8	9
header cell, first or last column	1	1

The flagged sentence lives in the header table under "While the cursor is on it", and the surface hands Tab back on a header cell before it ever looks at the column index — so one press is correct in its stated scope, now verified rather than assumed.

What was false is the page's opening summary: "one Tab enters it… and one more Tab leaves", unqualified. Under wrap-rows that is 8 and 12 on a twelve-cell grid. 12 is exactly rows × columns; backward from that same cell is 1, because it is already at the other release corner. The round-trip sentence was also wrong — Shift+Tab restores the cell the walk left from, not where it started.

Test Plan

• 132/132 browser tests across Chromium and WebKit on the integrated lane.
• Full unit suite, typecheck, typecheck:public, lint, prettier, build, api:check — clean.
• Mutation-proved throughout. Two worth noting: widening the movement union by a bogus member builds successfully and fails only the type test — catching drift the compiler does not; and restoring the wrap-rows corner clamp fails 10 of 18 exit tests while the four header tests per engine keep passing, since the header release is a separate branch.
• The rowModelDescriptor mutation catches the silent never case, which had no diagnostic before.

Found, not fixed — two are latent hazards

A scan turned up ~29 duplicated types. The diagnosed class is fixed; the rest are reported rather than sprawling the diff. Two are real hazards where a copy was silently widened:

• SurfaceFacade.getSnapshot().editing.status is string, not PretableEditStatus — and is compared against "editing" at pretable-surface.tsx:3273.
• csv.ts:348 takes {kind: string}, not PretableExpansionDefault — compared against "expanded", so renaming that literal would silently flip CSV completeness.

Also: PretableSortDirection has 8 byte-identical copies; PretableExportScope 10; ColumnType and PretableColumnType are byte-identical and both published from @pretable/core — two public names for one union.

🤖 Generated with Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

1 Skipped Deployment

Project	Deployment	Actions	Updated (UTC)
pretable	Ignored		Aug 17, 2026 3:34am

[github-actions]

Vercel preview ready

Preview: https://pretable-4sm0o0sa4-cacheplane.vercel.app
Commit: e39ee834ba07d71ba03f70dca7a07e1d7406bf9c

_{Updated automatically by the deploy-preview job.}