fix(row-model): ColumnValueOf resolves to unknown, not never

[blove]

The defect

value in rows-mode onRowChange was typed never. never is assignable to everything, so this guard — shipped in the editing page's worked example — compiled green while checking nothing:

if (columnId === "quantity" && typeof value === "number" && value < 0)

Root cause: ColumnValueOf fell back to never for any column that declares no accessor. No example in the docs corpus uses accessor columns, so this hit every rows-mode onRowChange the docs teach — not just this one page.

Found while investigating a report that the cell-editing page showed "invalid code". The file typechecked and linted clean; the defect was only visible by probing the type.

The fix

ColumnValueOf now falls back to unknown, which forces the guard instead of accepting it.

never is still load-bearing inside the distribution — it is the union identity, so non-matching members of a mixed column tuple must vanish rather than widen every answer. The fallback is applied outside the distribution via [X] extends [infer TResolved]; a naked extends infer would distribute, and distributing over never short-circuits the whole conditional to never, which is precisely the case the fallback exists to catch.

Both behaviors are pinned in column-value-of.test.ts — including the mixed-tuple precision assertions, which were already green before the fix and so will catch any widening.

Fallout, all fixed

• apps/bench passed columnId as never to distinctValues, forcing the id to match no column. BenchColumn declares an accessor, so plain columnId resolves correctly — the cast was papering over this same hole.
• grid/editing.mdx's value row claimed "inferred from columnId", true only for accessored columns. Replaced with an accurate note plus a narrowing snippet.
• The example passed <PretableSurface<StockItem>, which bought nothing (TRow is inferred from rows) and read as broken JSX.
• The docs-guard roster excused that table on the grounds that it documents PretableSurfaceProps.onCellEdit — a prop that exists nowhere in the repo. That stale excuse is why the false claim survived. Replaced with the real limitation: the member reader handles interfaces only and cannot bind a mapped-type-indexed union. Follow-up filed to teach it that shape.

Verification

Rebased onto origin/main (10 commits ahead, incl. #447 touching row-model); all gates re-run against the new base. The .api.md files merged cleanly and were regenerated to confirm the merged text matches what the generator produces.

• pnpm typecheck clean; pnpm lint clean (1 pre-existing unrelated TanStack Virtual warning)
• 2,073 package tests + 554 website tests pass
• api:check clean, with no ae-forgotten-export warning (this repo had zero; the fix was restructured to avoid introducing the first)
• Type-performance gate passes

One number worth a look: the extra conditional costs ~5% more type instantiations (deterministic: 126,366 → 132,161 on columns-500, against a 151,640 budget). Memory on columns-500 reads 98.9–101.6 MiB across runs against a 109.9 MiB budget — ~92% of ceiling, up from ~83%. Memory is noisy run-to-run; instantiations are not. I did not re-baseline the budgets, since that would silently loosen a gate.

🤖 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 3:30am

[github-actions]

Vercel preview ready

Preview: https://pretable-n089t4ye6-cacheplane.vercel.app
Commit: ae1f386d87407586247cd2966c5d0f465eca6d0a

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