Two independent clusters of default-behavior bugs from the beta triage.
Expanded and paginated state contents:
Expand-all materialized every id in `rowsById`, including leaf rows that
can never expand, so a serialized `ExpandedState` was polluted with dead
keys and `getExpandedDepth` was skewed by them. Materialization now writes
only ids where `row.getCanExpand()` is true, chosen over `subRows.length`
so `getRowCanExpand` lazy-load overrides stay expandable.
`getExpandedDepth` filters the same way, and `getIsAllRowsExpanded` now
only considers expandable rows so the materialized map still round-trips
as "all expanded" (a map of stale ids with no expandable rows is false).
Paginated `flatRows` pushed each row and then recursed into its subRows,
but page rows already contain expanded descendants inline, so those rows
appeared twice. The rebuild now dedupes by row id. Collapsed descendants
stay included, consistent with every other row model, where `flatRows`
ignores expansion state.
`row.toggleExpanded(bool)` and `table.toggleAllRowsExpanded(bool)` fired
`onExpandedChange` even when the requested state already matched, so
controlled consumers got spurious callbacks. Both now early-return.
Neither toggle consulted `row.getCanExpand()`, unlike their row-selection
equivalents, so a non-expandable row could be written into expanded state
imperatively. The expand direction is now guarded; collapsing is always
allowed so stale expanded ids can still be cleaned up.
Sorting toggle defaults:
`column.toggleSorting()` called `column.getNextSortingOrder(column)` with
no `multi` argument, so `enableMultiRemove` was dead on every path. The
argument is now forwarded as `multi && column.getCanMultiSort()`, matching
the multi-mode condition the updater itself uses. The public
`getNextSortingOrder` type now accepts the `multi` argument it always
supported at runtime.
`column.getAutoSortDir()` sampled only `flatRows[0]`, so a leading null or
a manual-sorting data swap flipped the inferred first direction mid-cycle
and silently dropped a state from the toggle cycle. It now samples the
first 10 rows for a non-nullish value, matching `getAutoSortFn`.
Docs: `sortUndefined: false` was documented as "considered tied" in all
framework sorting guides; undefined values are actually passed straight to
the sorting function with no special handling. Corrected, and the missing
`'first'`/`'last'` bullets were added to the generated reference.
Also fixes three crashes in the MRT examples surfaced while testing the
sorting changes. The MUI and Mantine sort labels read `sorting` off
`state`, which does not carry it, and crashed on `sorting.length`; they now
read it through `table.Subscribe` on `table.atoms.sorting`, which also
makes the badge reactive. The row-action edit handler passed a shallow row
copy to `setEditingRow`, breaking the row identity the edit modal needs.
`mrtRefsFeature` seeded the keyed `editInputRefs`/`filterInputRefs` bags
with `null` instead of `{}`, crashing on first keyed write. New e2e
coverage pins all three plus the sort-direction indicator.
Closes#6115Closes#5833Closes#6136Closes#4939Closes#4946Closes#5147Closes#5832
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>