feat(mosaic): per-column summary kind overrides for DataTableExplorer - #817

Problem

DataTableExplorer's columns option is all-or-nothing: a column is either loaded — with its histogram/category summary and cross-filter UI — or excluded from the table entirely. There is no way to:

• show a column's values while disabling its summary and filtering (e.g. Latitude/Longitude, where a histogram and brush add noise and query load but no insight), or
• override the Arrow-type-driven summary kind (e.g. a numeric code column that reads better as category buckets than a histogram).

The summary kind decision (histogram / category / unsupported) was hard-coded from the Arrow type in three separate places (useDataTableExplorerColumns, connectDataTableExplorerSummaryClients, createEmptySummaryState), so there was no seam to hook into. Hiding the summary UI from outside wouldn't help either: the summary Mosaic clients would still be created and still run their queries.

Solution

Add one deep option instead of a per-column config object:

<DataTableExplorer
  tableName={dataTable.table}
  selection={brush}
  getColumnKind={(field) => {
    if (field.name === 'Latitude' || field.name === 'Longitude') return 'none';
    return 'auto';
  }}
>

• resolveDataTableExplorerColumnKind(field, getColumnKind?) is now the single decision point for a column's kind; the three call sites consult it instead of re-deriving from the Arrow type.
• 'none' renders the column with no summary cell and creates no summary clients, so no summary queries run and the column doesn't participate in cross-filtering. Sorting and row rendering are unchanged.
• 'category' / 'histogram' force a specific summary. Incompatible requests (histogram on a string column, any summary on binary/geometry) fall back to the type-driven default rather than producing broken queries.
• Resolved kinds are value-memoized in useDataTableExplorer, so an inline resolver lambda can't tear down and reconnect every summary client on each render — the client lifecycle only restarts when a column's resolved kind actually changes.
• For custom headers built on useDataTableExplorer, the default summary cells are now exported: DataTableExplorerSummaryCell, DataTableExplorerHistogramSummaryCell, DataTableExplorerCategorySummaryCell, plus the resolver itself.

Design notes: columns stays a pure projection; render concerns (widths, labels, custom cells) stay in the compound-component layer. The resolver is a function, so persisted surfaces (dashboard panels / block documents) that want this can layer a serializable Record<string, kind> on top later — deliberately not included here. A read-only summary mode ("show histogram, disable brushing") is a distinct axis and also deliberately out of scope.

How to test

1. pnpm install && pnpm build
2. Unit tests for the resolver (default kinds, 'auto', 'none' always honored, incompatible overrides rejected):

   cd packages/mosaic && pnpm test

   24 suites / 136 tests pass, including the new resolveDataTableExplorerColumnKind block in __tests__/data-table-explorer.utils.test.ts.
3. See it live in the deckgl-mosaic example, which now sets Latitude/Longitude to 'none':

   pnpm dev deckgl-mosaic-example

   In the Earthquake Profiler table at the bottom:
   • Latitude / Longitude show name + type badge only — no histogram, no brush. Clicking the header still sorts.
   • Every other column is unchanged (DateTime/Magnitude histograms, MagType/Source category buckets).
   • Click the Md bucket under MagType: rows filter (≈45,330 of 54,936), all summaries update, and the 'none' columns stay summary-less.
   • Optional: temporarily return 'category' for Depth in EarthquakeProfiler.tsx's getColumnKind to see a numeric column forced to category buckets, with bucket-click filtering working.

All three steps verified locally (macOS, headless Chromium screenshots for step 3).

🤖 Generated with Claude Code