finding(fields): `DateTimeCellRenderer` and `DateCellRenderer` accept DIFFERENT `field.format` vocabularies — `format: 'relative'` is honoured on a `date` column and silently dropped on a `datetime` one

[claude]

Found while implementing objectui#8352 (the same defect class on the dataset-measure path) by the objectui domain:ui seat, session session_01611D6ZaRaMmwTNQmSbk8MH. Filed rather than fixed there: different package, different surface, and it needs its own verification surface. Not claimed, no code written.

The asymmetry

packages/fields/src/index.tsx, the two neighbouring cell renderers:

:919   DateCellRenderer      const style = dateField.format || 'relative';
:928                         formatDate(safe, style, { dueLike, locale, t });

:970   DateTimeCellRenderer  const style = field.format || 'compact';
:994                         formatDateTime(date, { style, locale, t });

Both read the author's field.format and hand it to a formatter. The formatters do not accept the same vocabulary:

• formatDate honours 'short' and 'relative'; anything else falls to the default locale face.
• formatDateTime honours 'compact' alone; anything else falls to the verbose default.

⇒ An author who writes format: 'relative' on a field gets the relative face on a date column and silently the verbose default on a datetime column. No error, no warning, no fallback — the same silent-drop signature as objectui#8352, one surface over. The reverse holds too: format: 'compact' is honoured on a datetime column and inert on a date one.

DateDisplayOptions.style's own docblock states the split as intended for the functions ("'compact' is formatDateTime's dense grid cell face, 'short' and 'relative' are formatDate's"). What is not stated anywhere is that the two CELL renderers therefore accept different authored vocabularies for the same key name on the same metadata surface.

Measured vs read

Measured (unit level, TZ=UTC, en-US, clock pinned to 2026-09-09T12:00:00Z, value 2026-09-11T09:30:00.000Z):

call	output
formatDateTime(v, { locale })	Sep 11, 2026, 09:30 AM
formatDateTime(v, { locale, style: 'relative' })	Sep 11, 2026, 09:30 AM — unchanged
formatDateTime(v, { locale, style: 'short' })	Sep 11, 2026, 09:30 AM — unchanged
formatDate('2026-09-11', 'relative', { locale })	In 2 days

Read from source, not driven: that DateTimeCellRenderer passes field.format into that style key unchanged (lines above). A browser run through a real datetime grid column has not been done, so the end-to-end claim is inference from a measured formatter plus a read pass-through — worth one driven confirmation before anyone acts on it.

Why it is a metadata trap, not a nitpick

field.format is authored metadata: written by one person, stored, and re-authored later by another. The runtime accepts the key, parses it, drops it, and renders something plausible. That is the shape that made objectui#8352 survive a triage pass — a reader cannot tell an honoured style from a dropped one by looking at the output, because both look like a legitimate date.

Scope note

Whether to fix this by widening formatDateTime's vocabulary, by mapping at the renderer, or by rejecting the unaccepted spelling is a real design call with a published signature nearby, so it is left to a ruling rather than guessed. objectui#8352 answered the analogous question on the measure path at the CALL SITE, mapping the vocabulary explicitly and refusing to widen the published signature; that precedent is available but was decided for a different surface and should not be assumed to transfer.

⛔ Not the same as objectui#8209 (the readonly datetime WIDGET faces, which do not call formatDateTime at all) and not objectui#8507's bare-em-dash family. Dedupe: one targeted search over the datetime / field.format / relative-style vocabulary returned 12 results including a known-present control, covering objectui#7443, objectui#7745, objectui#7816, objectui#8209 and objectui#8352 — none names this renderer-level vocabulary split. Not upgraded to exhaustive.

---

Generated by Claude Code