row_height_error_p95_px cannot fail for a nowrap grid, and the bench page reads meaning into it anyway

[blove]

Found while verifying #400.

row_height_error_p95_px is computed as |max(cell.scrollHeight) + padding − row.height| (apps/bench/src/bench-runtime.ts:1982-2001). A cell's scrollHeight is floored at its clientHeight, so with white-space: nowrap and overflow: hidden nothing can ever overflow vertically and the metric is structurally incapable of returning nonzero.

All three comparator adapters render exactly that way (see #400), so their committed 0 is not an achievement — it is a metric that cannot fail. This is the same shape as the estimator issue recorded earlier: a metric that never consults the thing it claims to measure.

It matters because the published page reads meaning into the one nonzero value. apps/website/app/bench/page.tsx:388-393 interprets AG Grid's 2px as "a sign that wrapped-cell layout doesn't round-trip through its line-height pipeline as cleanly as pretable's text-core does" — an inference about wrapped-cell layout, drawn about a grid that is not wrapping.

Two things to fix, and they are separable:

1. The metric should report not applicable rather than 0 when a grid renders no wrappable content, so a vacuous pass stops looking like a passing grade. Whatever the mechanism, it must be mutation-provable: show the metric failing for a grid that genuinely mislays a row.
2. The page's AG Grid inference should go, or be re-derived once S2 comparative scroll is not like-for-like: only pretable wraps #400 lands and the comparators actually wrap.

Note that #400's fix changes the input here: once the comparators wrap, this metric starts being able to fail for them, and the numbers it produces will be real for the first time. That is a reason to fix the interpretation on the page in the same pass, not before.

[blove]

Scoped part 1 but did not start it, to avoid leaving a half-finished typed change. Mapping it here so it is a pickup rather than a re-investigation.

Detection. The reason the metric cannot fail is that scrollHeight is floored at clientHeight, which is exactly what white-space: nowrap guarantees. So the applicability test is a direct read of the thing that causes it:

const nonWrapping = new Set(["nowrap", "pre"]);
const wrappable = cells.filter(
  (cell) => !nonWrapping.has(getComputedStyle(cell).whiteSpace),
);
if (wrappable.length === 0) return undefined;  // not applicable, not zero

Values that DO wrap and must stay measured: normal, pre-wrap, pre-line, break-spaces.

Call sites that change with the return type (apps/bench/src/bench-runtime.ts):

• :1982 measureWrappedCellRowHeightError — returns number | undefined
• :1817 VisibleRowSample.heightError — widens to number | undefined
• :1843 profile.measureRowHeightError(row, rect.height) — assignment site
• :498 and :785 — both already .filter((value) => value > 0), which drops undefined today by coercion. They should filter on !== undefined explicitly, or a not-applicable row and a genuinely-zero row stay indistinguishable, which is the bug one layer down.

With no wrappable rows the array is empty and percentile([]) omits the metric, so the artifact carries no row_height_error_p95_px at all rather than a 0. That is the honest shape: absent, not passing.

Mutation proof required by this issue. A test that mislays a row — render a wrapping cell whose content exceeds its row and assert the metric reports the difference — plus its negative: the same fixture with white-space: nowrap must yield undefined, not 0. Without the second half the guard is the same vacuous pass it replaces.

Sequencing. Part 2 (the bench page's AG Grid inference at apps/website/app/bench/page.tsx:388-393) should ride with #400/#415 as this issue says — once the comparators wrap, that number becomes real for the first time and the sentence can either be re-derived or dropped on evidence. Doing it before would just be deleting a claim someone then has to rewrite.