--- id: "packages/active-grid/changelog" title: "Changelog" description: "What changed for users of active-grid, per version." url: "/packages/active-grid/changelog" locale: "en" revision: "dbc0cc89a4f476a0" --- # Changelog > [!NOTE] > **Release state** > > The version line at the top of this page is read from the package and the > release register of Foundation. A recorded release counts only for exactly the > content it recorded: when the package changed after it, this site does not > call that version released. The section Unreleased lists what changed since, > including the changes made since Foundation maintains the package. Notable changes to `@workspace/active-grid`. Format based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/); versioning per [Semantic Versioning](https://semver.org/spec/v2.0.0.html). For architectural rationale of larger changes see `docs/DECISIONS.md`. --- ## \[Unreleased\] ### Maintained in Foundation (since 2026-10-21) - **Dependencies: `@workspace/active-helpers` replaces `@workspace/active-utils`.** The serialization, export-font and menu helpers the grid uses (`./serialization`, `./export-fonts`, `./menu-items`) now come from `@workspace/active-helpers`, the one package that owns them, instead of the `portable-helpers` selection of `@workspace/active-utils`. Same functions, same behaviour; the grid's public API is unchanged. A host that receives the grid receives `active-helpers` with it (the transfer closure names it). - **Fixed: a wide grid no longer widens the page.** `.grid-root` is now `position: relative`, the containing block of the grid's absolutely positioned parts (visually hidden texts, the skip-link target, the announcer). Before, those took a containing block outside the grid: a grid wider than its host's scroll container (a phone, a narrow column) made the whole page scroll sideways. Found on the documentation site at 390 px. - **Tests run without the source repository.** The source contracts (`candidate-boundary`, `consumer-precheck`, `runtime-reach`) read through a package-local helper (`src/__test-utils__/files-if-present.ts`) instead of the host's `repo-tools`; one test sorts with a local machine-order comparator instead of `@workspace/active-utils/sorting`. The scrollbar test checks the grid's own thumb tokens and no longer reads one host's brand mapping of them (that mapping is the host's test). - **Documentation:** references to documents of the source repository are text, marked `(source repository: …)`, instead of links that resolve nowhere here. ### Changed — BREAKING for the managed package set (ADR-059) - **Formatting comes from `@workspace/active-format`**, not from the `formatters` entry of a translation package. The manifest declares `@workspace/active-format` instead of `@workspace/active-i18n`; the package imports nothing from a translation package any more. No prop, export or behaviour of active-grid changed: its presentation type is still the `ValuePresentation` union, now owned by `@workspace/active-format`. - A consumer that formats next to active-grid with the same functions imports them from `@workspace/active-format`. --- ## \[11.0.0\] - 2026-10-18 - The canonical transfer form, and the first recorded release ### Changed — canonical transfer form (2026-10-18, platform ADR-112) - **The manifest is the transferable manifest** (RPS-D7): no workspace package in `devDependencies` (what the package's own development takes from its host repository, such as the lint preset, is host tooling in the reusable package register); the declaration packages of what the source imports sit in the field of their module (`@types/react` and `@types/react-dom` as optional peers beside the `react` peers). Nothing changes for an existing consumer. - **`tsconfig.json` is self-contained**: it no longer extends the host repository's shared TypeScript preset, with the same compiler options, because a bundler in a consumer reads it for every source file it compiles. ### Changed (2026-10-04, one menu pipeline and behaviour with active-charts, ADR-058) - **Added (G-1):** custom context-menu items nest to any depth; a `children` inside `children` is a submenu of the submenu. - **Added (G-2):** a disabled item's `disabledReason` is exposed as `aria-description` (it was a `title` only), on submenu triggers too. - **Changed (G-3):** the resolve pipeline is `resolveMenuItems` from `@workspace/active-utils/menu-items`, shared with Active Charts; same semantics, the permission check runs as its `keep` step. - **Changed (keyboard):** a menu opened by Shift+F10 or the ContextMenu key focuses its first item; Tab closes the menu (not during IME composition). - **Changed (touch):** under a coarse pointer every menu item is at least 44 px high. ### Changed (2026-10-04, a declared menu permission fails closed) - **Behaviour change:** a context-menu item that declares `permission` is no longer shown when no `hasPermission` checker answered it. Before, a menu built without a checker (or passed directly as a `ContextMenuBuilder`) ignored `permission` and showed the item to every viewer. Now `createContextMenu` without a checker leaves such an item out, the render boundary keeps an item with `permission` only when the factory approved it, and it reports a dropped item through the `` sink (`warn`, source `ActiveGrid.ContextMenu`, `data: {label, permission}`). The factory returns approved items as fresh copies; the consumer's objects are never mutated. **Migration:** a menu that declares rights passes the viewer's checker: `createContextMenu({hasPermission, customItems})`. The report reaches the sink outside production only, like the other configuration warnings; leaving the item out does not depend on it. ### Changed (diagnostics family parity, R-58, ADR-057, 2026-10-02) - **Every diagnostic reaches the `GridDefaults` sink.** The thirty configuration warnings that wrote to `console.warn` directly now go through `` (source `ActiveGrid`), still outside production only. Without a sink they reach the console as before, as `[ActiveGrid] …`. - The currency, percent and calendar-day preset warnings are reported once per column when the columns are defined, no longer once per rendered cell. The CSV separator and the capability-drift warning of `createResourceServerFetchFn` are reported once per ground; the drift carries both capability maps as structured `data`. - Messages no longer repeat an `[active-grid]` or `[ActiveGrid]` prefix. - No public API change. ### Fixed (re-convergence, 2026-09-30) - **Server mode: a re-render no longer restarts the adjacent-block prefetch.** The prefetch effect depended on a page object the block cache creates on every render. Every host re-render therefore aborted and reissued the request for the next block. - **Server grouping: a load-more that settles after a sort, filter or grouping change is dropped.** Before, its rows were appended to the group in the new dataset, and a failure set the grid's error. Only the first page of a group was checked against the dataset. - **Undo: a failed undo keeps what was already undone redoable.** The failure path re-recorded the entry, which cleared the whole redo stack. - **A select filter restored from the URL matches again.** The URL brings a numeric-looking option code back as a number, and the select predicate compared it with the cell's written form. Numbers are now compared by their written form; other non-strings still never match (the string-key contract of `filterOptions`). - **Group labels are isolated bidi runs.** The client group label, the server group value and the grouping chip render `dir="auto"` with `.grid-isolate` (ADR-052 point 4). - **The lookup dialog's hint has its style again.** It is small, muted text in the footer. ### Changed (portable runtime P9, ADR-056) - **`@date-fns/tz` is a declared dependency.** The inlined formatter closure imports it; an install outside the workspace reads only this manifest, so the package now declares it. No runtime change. - **Documentation.** README, functional overview and the `locale` prop describe the standalone product (no Next, no active-ui, the stylesheet import, the time-zone requirement of date filters and calendars, optional export backends as bundler externals). ### Documentation (portable runtime P8D, ADR-055) - **Export guarantees.** `docs/guides/export-guarantees.md` states the tiers L0 to L3b-print per format and script, the route to take, the error kinds, a host font example and the host L3b recipes. No runtime change. ### Changed (portable runtime P8C, ADR-054) - **BREAKING (behaviour): the PDF refuses what it cannot draw.** Without an export font set, text outside WinAnsi fails with `GridExportError` `EXPORT_FONT_REQUIRED` instead of drawing broken glyphs; complex and right-to-left scripts, residual combining marks, joiners and variation selectors, emoji and supplementary-plane characters fail with `EXPORT_TEXT_UNSUPPORTED` whatever the fonts. WinAnsi-only PDFs without a set are unchanged. - **`xlsx`, `jspdf` and `jspdf-autotable` are optional dependencies.** The package root no longer reaches them; a missing one fails its export with `EXPORT_DEPENDENCY_MISSING` (`missingModule`). ### Added (portable runtime P8C, ADR-054) - **Export fonts for the PDF.** `GridExportFonts` (with `GridExportFontFamily`, `GridExportFontFace`, `GridExportFontSource`): host families in fallback order, faces as bytes or lazy loaders. Accepted by `GridPdfExportOptions.fonts`, the grid's `export.fonts` and `GridDefaults`' `export.fonts`, in that precedence; `'inherit'` uses no set. Each cell draws in the first family that covers all of it (`EXPORT_GLYPH_COVERAGE` otherwise); the flat, grouped and server-grouped PDF share one pipeline. Structurally the charts' `ChartExportFonts`. - **`GridExportError`** (`GridExportErrorCode`, `GridExportErrorOptions`): stable codes, metadata without exported text. - **Export direction.** `GridExportDirection`; `GridPdfExportOptions.direction` puts the first column on the right; `GridExcelExportOptions.direction` opens the sheet right to left (`Workbook.Views[0].RTL`). The row menu passes the grid's resolved direction. Values stay logical and typed. ### Fixed (portable runtime P8C, ADR-054) - **The PDF export works in the browser.** It called `doc.autoTable(...)`, which jspdf-autotable 5 installs only on a global `window.jsPDF` that the ES build of jsPDF never sets (measured: `TypeError: doc.autoTable is not a function`). It now calls `autoTable(doc, options)`. ### Changed (portable runtime P6, BW-3, ADR-053) - **BREAKING: pinning is logical.** `ColumnPinningPosition` is `false | 'start' | 'end'`, `ColumnPinningState` is `{start, end}`, `column.pin('start' | 'end')`; the readers are `getStart*`/`getEnd*` (`getStartVisibleLeafColumns`, `getEndHeaderGroups`, …) and header group names `start_*`/`end_*`. No alias for `left`/`right`. - **BREAKING: `columnResizeDirection` is removed.** The grid's one resolved direction reaches the options as `direction`. - **BREAKING: persisted state version 3.** A version 2 payload is cleared on load; a physical `left`/`right` pinning slice is dropped. - **RTL.** `dir="rtl"` (or an inherited rtl) lays the grid out mirrored: pinning, scrolling, column virtualization (active-virtual reads the rtl offset), header/body sync (TD-14 fixed), resize, reorder, keyboard (ArrowLeft/Right visual, Home/End logical), selection, menus (submenus toward the inline end), filters and the calendar; directional icons mirror. `cellAlign: 'right'` is the inline end. - **Formatted values read in their locale's direction.** A value renders with its presentation locale's script direction where that differs from the grid's (`GridValue`), so an ltr-locale date keeps its order in an rtl grid. ### Added (portable runtime P5, ADR-052) - **`dir` prop.** A grid resolves one UI direction: the explicit `dir`, else the page's direction (the nearest `dir` attribute, followed when it changes), else `ltr`. Never the locale. Every popup the grid opens carries it, and Base UI's keyboard semantics follow it. Until P6 a grid that resolves `rtl` throws `GridProjectionFault('unsupported-direction')`; an invalid `dir` throws `GridConfigurationError` (`GRID_DIRECTION_INVALID`). - **`meta.textDirection`** (`GridTextDirection`: `auto`, `ltr`, `rtl`). Cell text, headers, filter options and the cell tooltip render as isolated runs: free text `auto`, formatted values inheriting, codes as declared. A right-to-left column name in an `aria-label` or announcement is isolated with FSI/PDI; Latin labels in an `ltr` grid are unchanged. Exports and the clipboard carry no bidi controls. ### Changed (portable runtime P4, BW-2, ADR-051) - **BREAKING: an unsupported locale throws.** A grid whose presentation names a tag the formatters cannot honour (malformed, without Intl data, a non-Gregorian `-u-ca-` or an unknown `-u-nu-`) fails with `UnsupportedLocaleError` from `@workspace/active-i18n/formatters` instead of rendering English. Any other BCP 47 tag Intl supports formats in its own convention and digits. - **The calendar accepts every supported locale.** Month and weekday names in the Gregorian calendar, accessible labels from Intl ("Next month", "today"), day numbers in the locale's digits, and the week start from `Intl.Locale` week info (Monday without it). A `calendarLocale` object that carries its own `labels` keeps them. The previous five-language refusal is gone. - **Number input reads any supported locale.** The number editor and the numeric filters use the one portable parser: the locale's decimal separator, Indian and other grouping on paste, every Unicode digit set and the U+2212 minus. Parentheses and a trailing minus stay unsupported. - **IME composition is left to the input method.** While an input method composes, Enter, Escape, Tab and the arrows neither commit, cancel nor move the active cell. ### Changed (portable runtime P3, BW-1, ADR-049, ADR-050) - **BREAKING: the grid ships its own stylesheet.** Import `@workspace/active-grid/styles.css` once, after declaring the cascade layer order (`… base, active-grid, components, utilities`). All rules are plain CSS in `@layer active-grid`; no Tailwind scan and no design-system stylesheet is needed. A neutral light and dark baseline is built in; a host brands the grid by mapping the `--grid-*` role tokens (`--grid-surface`, `--grid-primary`, `--grid-border`, `--grid-ring`, `--grid-status-`, the radius and type scale, …). Previously unlayered, the grid's rules now lose to a host's utility layer, so a host class on a grid element takes effect. - **BREAKING: `StatusTone` → `GridStatusTone`** (same five values) for `meta.statusTones` and ``. The grid no longer exposes any type from `@workspace/active-ui`; `FilterPopoverContent` takes the grid's own popover props. - **BREAKING: `LinkCell` renders a native ``** for internal links. A host with a client router passes its link once via `` (`GridLinkComponent`, `GridLinkProps`: `{href, className, children}`). External links are unchanged. - **BREAKING: default menu classes.** `ui.contextMenu` and `ui.headerMenu` default to the package classes `grid-menu-popup--context` / `grid-menu-item--context` and `grid-menu-popup--header` / `grid-menu-item--header` (same sizes as before). - The grid renders on an internal primitive layer (Base UI and react-day-picker); its markup names only `grid-*` classes. - **Dependencies:** `@workspace/active-ui` and `tailwind-merge` removed, `react-day-picker` added; the `next` peer removed. Peers are `react` and `react-dom` only. ### Changed (portable runtime P1, ADR-047) - **Diagnostics go to a host sink.** `` takes a `GridDiagnosticSink` (`{level: 'debug' | 'warn' | 'error', source, message, data?}`, a caught error under `data.error`); nearest boundary wins. Without one, `warn` and `error` go to the console and `debug` is dropped. New public types `GridDiagnostic`, `GridDiagnosticLevel`, `GridDiagnosticSink`. `copyToClipboard` options take `diagnostics` too. - **BREAKING (internal consumers only):** a misconfiguration throws the new, public `GridConfigurationError` (`code` `GRID_ROW_ID_MISSING` or `GRID_SERVER_OPTIONS_INVALID`, `field`, `constraint`, `value`) instead of `ValidationError` from `@workspace/active-errors`. Messages are unchanged. - `CellErrorBoundary` is a function component over an internal class boundary; its props and rendering are unchanged. - **Dependencies:** `@workspace/active-logging` and `@workspace/active-errors` removed. The CSV/decimal helpers come from `@workspace/active-utils/serialization` instead of `/helpers`. ### Added (FAM-V2-T6, ADR-046) - **The outcome of a menu export reaches the host.** `GridExportOptions.onExportComplete(format, scope)` fires after the file was handed to the browser; `onExportError(format, error, scope)` when no file was produced. Before, a failed menu export was only logged and the host could not tell. Without `onExportError` the failure is still logged, never swallowed. New type `GridExportScope` (`'page' | 'loaded' | 'selection'`). Additive; the programmatic export functions keep reporting by throw, rejection and return. ### Breaking (FAM-V2-T4, ADR-045) - **`GridDefaults` replaces `GridConfigProvider`.** One public defaults boundary carries the presentation (`locale`), the translator (`t`) and the `PartialGridConfig` sections as top-level props. Migration: `` becomes ``. Nested boundaries refine the enclosing one; presentation and translator are replaced whole, sections merge one level deep, `filtering.operatorPresets` per name. - **No implicit English.** A grid without a `t` prop and without an inherited `` throws a configuration error. The English table is the exported `englishGridTranslator`; pass it on purpose for standalone use, fixtures and examples. The operator factories (`getTextOperators`, …, `getOperatorsForType`) take a required translator, and the English-baked `TEXT_OPERATORS`, `NUMBER_OPERATORS`, `DATE_OPERATORS` and `BOOLEAN_OPERATORS` are removed. - **Built-in column types are the package default.** Every grid applies them without a prop; `columnTypes` extends or overrides per type name. `getDefaultColumnTypes` leaves the public API: remove `columnTypes={getDefaultColumnTypes()}` and spread nothing into a custom map. - **Provider surface narrowed.** `GridTranslationProvider`, `useGridCalendarLocale`, `useGridConfig`, `DEFAULT_GRID_CONFIG`, `GridProvider`, `FlexLayoutContext` and `FlexLayoutContextType` are internal. `useGridTranslation` and `useGridLocale` stay public and return the effective values of the enclosing grid. - **JavaScript theme layer removed** (FAM-W6): the `theme` prop, `createGridTheme`, `GridTheme`, `gridThemes`, `GridThemePreset`, `compactTheme`, `comfortableTheme` and `financialTheme`. The grid CSS stays in active-ui; a scoped override sets a `--grid-*` value on `.grid-root` through `className`. Row height has one authority (`settings.rowHeight`, then `layout.rowHeight`). The per-grid scrollbar thickness setting went with it; the lane is 10px. ### Changed - `locale` is an optional instance override on `ActiveGrid` and `LookupGrid`: the prop, else the nearest ``, else a configuration error. `t` on `LookupGrid` follows the same rule. - `GridPresentation` is an alias of the active-i18n `ValuePresentation` (active-i18n D23); the type is unchanged. - The filter builder takes a condition's number and date in the grid's own controls instead of native `number` and `date` inputs. A number is typed in the notation of the grid's presentation, as in the number filter, and stored as a number; a date is typed in the presentation's date notation or picked from a calendar behind its own button, and stored as an ISO calendar day. Half-typed or unreadable input leaves the condition's value as it was, and emptying the field makes the value absent (it was `''`, which a number operator read as `0`). The operator stays a select, as in the column filters. - The pagination footer counts selected rows without building a row view per selected row. The displayed count is unchanged. - Row and column virtualization run on `@workspace/active-virtual` 2.0.0 (`useVirtualAxis`). Observable behavior is kept: the same windows, overscan, scroll-to-cell behavior and column resize; the measured row heights are now reset before the browser paints after a change of the row model, and the whole scroll range stays reachable for keyboard navigation to the last row when pinned rows or a footer follow the rows. No public API of the grid changes. - Diagnostic severity follows the repository meaning of ERROR and WARN. Recovered or expected conditions now log at WARN instead of ERROR: a row that is gone when an edit starts, commits or is buffered in batch mode, an edit cancelled because its row left the data, a refused persistence write (quota, private mode) and a refused clipboard write (permission; the function still returns `false`). Commit and batch-commit failures, export, render and fetch failures, and the refusal of a group row stay at ERROR. Messages are unchanged. ### Fixed - An incomplete filter-builder condition no longer filters rows. `evaluateFilterExpression` ignores a condition until its operator has every value it reads, as its documentation already said: adding a number condition before typing its value used to hide every row. `0` and `false` still filter, a range needs both ends, and `isEmpty` / `isNotEmpty` stay active without a value. An AND or OR group judges only its complete conditions (an incomplete range inside an OR group used to match every row), and a group with none matches every row. - The filter-builder toolbar button counts only conditions that filter rows, on the same rule as the evaluator. A condition still waiting for its value no longer shows a count or marks the button active; `isEmpty` / `isNotEmpty` count as soon as they are chosen. The panel's own count still covers every condition in it. - When row virtualization switches on while the reader is scrolled (rows loaded past `virtualizationThreshold`, or client data replaced by a larger array), the record at the top of the viewport stays there, at the same distance into it. The two bodies can lay rows out at different heights, so an unchanged `scrollTop` showed another record, up to many rows away. A grid at the top stays at the top; horizontal scroll is untouched. - Undo, redo and the edit history act only while the grid is mounted. An undo or redo whose `onDataCommit` finishes after the grid was unmounted no longer patches the old grid's rows, calls the consumer's `getRowId` or rewrites its history, also when the call is rejected; an ordinary commit finishing after unmount records no history. The consumer's write itself still completes. - A client group row's label cell no longer carries `data-wrap="true"` from a `wrapText` column: the label stays on one line, as on a server group row. A label too wide for its column now ends in an ellipsis with its full text as `title` (client and server), and the item count no longer wraps below it. - `wrapText` with the default text renderer now actually wraps. The renderer truncated by default whatever the column said, so a `wrapText` column, including one with `autoHeight`, still showed one line with an ellipsis and its row never grew. With `truncate` unset, a `wrapText` column now renders untruncated text on every default text path (plain, `valueFormatter` text, number, currency, percent and date semantics, and the empty-value placeholder), so the text breaks over lines and `autoHeight` rows grow to hold it. An explicit `truncate: true` still truncates, an explicit `truncate: false` still does not, columns without `wrapText` are unchanged, and custom `cell` renderers are untouched. - Under row virtualization, a toolbar search now resets the measured row heights, as sorting, column filters, grouping and expansion already did. A search puts other records at the row positions; an auto-height row kept the height measured for the record that was there before. - A measured row under row virtualization now keeps its own height when rows are pinned to the top. The row reported its display position (pinned rows first) to the row virtualizer, which counts only the unpinned rows, so every measured height landed on the row that many positions further down: an auto-height row was clipped to the estimate while its neighbour grew, and the last row's measurement was dropped. Server group and load-more rows had the same path. Keyboard navigation, cell ranges and `aria-rowindex` keep the display position. - Ctrl/Cmd+Shift+Z now redoes, and Ctrl/Cmd+Z and Ctrl/Cmd+Y also work with Caps Lock on. The handler compared the letter in lower case only, while a browser reports it in upper case under Shift or Caps Lock. - Documentation matches the source again: `enableUndoRedo` defaults to `true` when `onDataCommit` is set, records only single-column commits and keeps its history per mounted grid; the cell editors are the grid's own number input and a calendar-backed date input, not native `type="number"`/`type="date"` fields; `gridRef.removeRow` is a post-commit patch, not an optimistic update; the exclusive-selection recipe states that the consumer translates the `SelectionQuery` itself; the header checkbox is three-state for assistive technology and two-state on screen. The `calendarLocale` docblocks no longer show a locale the calendar refuses. Consumer guides and API pages that were partly Dutch are English. - A "Save all" that completes while the user kept editing no longer erases the newer edits (CR2-02). Its completion cleared the whole dirty buffer: a cell changed again after the save began was marked clean although it was never sent, and when `onBatchCommit` returned rows the returned (older) value replaced the newer one on screen. The completion now settles only the cells that save sent. A cell edited again, a cell first edited meanwhile and a cell edited back to its old value stay dirty with the value on screen, keep "Save all" available, and go out with the next save; a returned row no longer overwrites a cell that is still dirty. A rejected save keeps every change, and a second "Save all" while one is pending sends nothing. With no edit during the save, nothing changes. - Unmounting the grid ends a pending edit (CR-01). A validator that settles after the grid unmounted no longer leads to an `onDataCommit` call, in cell and row mode, and a pending batch flush no longer fills the buffer. A write the consumer already received finishes; the grid then patches no row and opens no editor. The same holds for "Save all". - A write that finishes after the user opened another cell no longer replaces that edit (CR-02). Enter, Tab and a `commitOnBlur` switch opened the next cell as soon as the earlier write completed, closing the newer editor and discarding what was typed in it. They now move on only when the committed edit is still the one in progress; the newer editor keeps its exact draft. Normal Enter and Tab progression, and a refused commit that stays put, are unchanged. - `aria-rowcount` and the body's `aria-rowindex` no longer count a filter row that is not rendered (R-54). A grid with one header row and no filter row announced its first body row as row 3 and one row too many in total; it now announces row 2. Grouped header levels, pinned rows, the virtual body and the group and "load more" rows of server grouping follow the same count. With a rendered filter row nothing changes. - An edit that ends while its validation is pending no longer commits (IR-01). Removing the row, cancelling the edit, starting another edit or "Discard all" in batch mode now ends the pending commit: `onDataCommit` is not called and nothing enters the batch buffer when the validator settles later. A validator's late message is no longer written under a cell opened afterwards. A live edit still commits exactly once. With `commitOnBlur`, moving to another cell commits the open edit first and opens the next one once that commit is accepted; a refused value keeps the open edit. - A grid with multi-row selection declares `aria-multiselectable="true"` also without cell selection (IR-02). Assistive technology was told the grid could not select several rows while it could. - The header checkbox announces a partly selected page as `aria-checked="mixed"` instead of checked (IR-03). Its look and what a click on it does are unchanged. - Turning `enableCellSelection` off ends every range, reported once as `[]` (IR-04). The range stayed stored out of sight: it came back when the feature was turned on again, and while off it still took the Ctrl+C hint from the selected-rows item in the context menu. - Server group rows carry `aria-rowindex` like every other body row (IR-05), so group, member and load-more rows form one unbroken sequence, also in a virtualized body. - `createResourceServerFetchFn`'s `onError` also fires for a read that fails by a rejecting or throwing action or a throwing `transform` (IR-06), once, after the retries are spent. It fired only for `{success: false}`. Redeemed and superseded attempts still notify nobody. - The server-grouping "load more" row works without a pointer (R-44). Its cells are cells of the keyboard navigation: the arrow keys reach the row, and Enter or Space on any of its cells loads the group's next block once, the same action a click takes. The row is named by its label, its cells announce the keys (`aria-keyshortcuts`), and it is `aria-busy` while the block loads; keys pressed in that time ask for nothing more. When the block lands, the active cell keeps its position, which then holds the first new row, and the focus is there. Before, Arrow Down moved the grid's position onto the row while the focus and the mark stayed on the row above, and Enter loaded from a row the reader could not see was active. - Focus that a replaced cell held returns to the active cell in every browser. The recovery relied on the `focusout` a removed element dispatches, which not every engine does; a cell that loses its registration while it holds the focus now takes the same path. - `pagination={{enabled: false}}` no longer pages a client grid (R-51). The pager was hidden while the rows were still cut to the page size, so a grid of 500 rows showed ten and the rest were out of reach. Every filtered and sorted row is now presented, a page size is ignored, and row virtualization still limits what is mounted; `aria-rowindex` counts from one. Server grids are unchanged. Consumer grids with pagination off now show all their rows. - A column width set by dragging survives a reload and every later viewport resize (TD-50 b). The flex layout ran before the stored column locks were read and wrote its own width over the stored one; no layout pass now runs before the locks are known. - `createResourceServerFetchFn`'s `onError` fires once per failed read, after the grid's retries are spent (TD-50 f). It fired for every attempt: one failure gave up to four toasts, and a read a retry redeemed still gave one. Retries and the grid's own diagnostics are unchanged. - A second click on the active cell changes nothing. It used to drop the cell's ring while its corner cue stayed, a half-drawn state. - Escape never leaves half a selection on screen: with a range it ends the range and leaves exactly the look of a click on the active cell; without one it changes nothing. It used to leave the ring and remove only the cue. - The first click after a reload no longer draws its mark bright and then dim. The row under the pointer faded its hover background in over 100 ms, and a browser applies no hover after a reload until the first pointer event, so that fade started under a mark that had just been drawn. The hover background now changes at once. - After a drag, the focus and the grid's active cell are the same cell (the drag's start). The press focused the start cell while the active cell stayed where it was, so the next arrow key moved from a cell the reader had left. - With rows selected, a click on a cell (including on a row's checkbox) no longer takes Ctrl/Cmd+C from the rows. The click left a one-cell range, which went before the selected rows and, on the checkbox column, copied nothing. - A right-click on a body cell outside a range makes that cell the active cell. It already took focus natively while the grid kept its previous active cell, so after the menu closed the next arrow key continued from a cell the reader had left. - `rowMenu.enabled: false` is honoured: no row offers the grid's context menu. The setting was declared and documented but not read. - The row menu's Copy copies a pinned row whose data position is on another page. It looked the row up in the current page and wrote nothing. - Keyboard operations on a pinned row act on the row shown there (R-47). They used to act on the row at that index of the data order: - Enter calls `onCellActivate`, or starts the edit, for the pinned row, and for the centre rows next to it. - ArrowLeft/ArrowRight on a server group row expands or collapses the group that has focus. It no longer toggles the group at a pinned member row's data index. - Arrow navigation past a column span uses the span of the row shown. - Cell-range selection and copy (R-45), with `enableCellSelection`: - Shift+click extends the selection from the active cell, which stays the anchor, to the clicked cell. It used to select only the clicked cell. - A range copy and a single-cell copy (Ctrl/Cmd+C) now go through the same clipboard TSV boundary as the row copy: a value a spreadsheet would run as a formula is neutralized with a leading apostrophe, and a value with a tab, line break or quote is quoted. The copied text is still the presentation form. - A range ends when a sort, filter, row pin or column move, hide or reorder puts other rows or columns at the positions it covers; one that leaves those positions in place keeps it. A range used to survive such a change and copy the data that had moved in. It no longer ends merely because the row count changed. - Ctrl/Cmd+C reports its outcome through `meta.onCopySuccess` and `meta.onCopyError`, once per copy, like the row menu's copy. The count is the number of rows copied: the selected rows, the rows a cell range spans, or 1 for a single cell. - A range or active cell on a pinned row copies the row shown there; it used to copy the row at that index of the data order. - Ctrl/Cmd+C copies a range made with the mouse alone; it used to do nothing. - Clearing a number filter right after an operator change clears it; the still-debounced value no longer re-applies the filter under the default operator (TD-50). An operator change re-applies the typed value as a number, not as the input's string. - A column drop lands on the side of the target header where the pointer is released. It used to read the side the last `dragover` had written to React state, which a drop in the same frame could see before it rendered (TD-50). - `unpinRow` refuses a missing row identity, as the kernel's `unpinRow` does, instead of filtering the pinned buffer by an absent identity (TD-46). - An empty grid is valid for assistive technology: the "No results" message is one row with one cell, and a body still loading its first rows carries `aria-busy="true"`. A row group without a row was a critical `aria-required-children` violation (R-27). - The focus marks follow the configured accent theme (UI-F1, UI-F2). The outer grid element carries the class `grid-root`, on which active-ui re-declares every theme-derived grid token, so a focused filter field no longer shows the default blue under another accent. The active body cell is marked at 50% of `--ring`, the active-ui field focus, instead of the full ring. Focus behaviour is unchanged. - The grid container's keyboard focus is subdued (UI-F3). Its tab stop marked itself with a full `border-primary` and a `ring-2` halo; it now carries its own border at 50% of `--ring`, no ring and no outline, the same expression as the active cell and the filter fields. Focus behaviour is unchanged. - Keyboard focus stays in the grid when the active cell leaves the DOM (R-27). Scrolling a virtualized row or column out of the window, or a render that replaces a control inside a cell, used to drop focus to ``, and the grid then ignored the arrow keys on its own tab stop. Focus now parks on the grid, returns to the active cell when it mounts again, and the next navigation key resumes from the active cell (clamped to the current model). - Toggling a row selection or an expander with the keyboard keeps the focus on that control. The system columns' renderers keep their identity when the consumer passes new column definitions, so their cells no longer remount. - Action mode no longer stops on a control's hidden native proxy (the `aria-hidden` input of a checkbox). ### Added - The body context menu names every copy action by its scope. In a range context: "Copy selection" and "Copy selection with headers". In a cell context: "Copy cell", "Copy row", "Copy row with headers" and, while rows are selected, "Copy selected rows (N)". In a row context (no cell selection): the selected rows, or the row, each with and without headers. "Copy selection" writes exactly the block Ctrl+C writes; "with headers" adds one header row with the columns' labels, also over several ranges. - The Ctrl+C hint sits on the one menu item that does what Ctrl/Cmd+C does at that moment, exposed as `aria-keyshortcuts`; where no item matches, none carries it. - `ContextMenuBuilder` receives an optional second argument, `ContextMenuContext` (new export): `kind` (`range`, `cell` or `row`), `columnId`, `selection` (`rangeCount`, `cellCount`) for a range, and `selectedRowCount`. A one-argument builder keeps working unchanged. - With cell selection, a client group row offers the grid's copy-only menu: "Copy cell" on a data column, or the selection items inside a range. The row builder, edit and pin never appear there. - Cells in a range carry `aria-selected="true"`, and a grid with cell selection declares `aria-multiselectable`. - New translation keys, supplied by the host: `grid.copy.selection`, `grid.copy.selectionWithHeaders`, `grid.copy.cell`, `grid.copy.row`, `grid.copy.rowWithHeaders`, `grid.copy.selectedRows`, `grid.copy.selectedRowsWithHeaders`. - Shift+F10 and the ContextMenu key open the row context menu for the active cell, with the same target as a right-click on it. The grid opens the menu itself; the browser's own `contextmenu` for that keystroke is suppressed. Escape closes the menu and focus returns to the active cell. - With `enableCellSelection`, the body context menu resolves its target when it opens: a right-click inside a range of more than one cell keeps every range and the active cell where they are (range context); a right-click outside all ranges clears them and makes the clicked cell active (cell context). Row selection is not changed, and the row's own menu items stay. - `useKeyboardNavigation`'s cell props carry `onMouseDown`: a primary-button press makes the cell active on the press itself. `onClick` stays and changes nothing after that press. - Styling hooks for the selection look, all set by the grid: on the grid element `data-cell-selection` (cell selection is on), `data-cell-range` (an explicit range exists) and `data-body-focus` (the focus is in the grid's body or in its own body menu); on a cell `data-cell-cue` (the cell that carries the selection's cue). Two tokens on `.grid-root`: `--grid-range-outline` and `--grid-range-fill`. - While the focus is parked on the grid (the active cell scrolled out of the mounted window), Shift+arrow keeps extending the range, Ctrl/Cmd+C copies, Escape ends the range and brings the active cell back, and Shift+F10 or the ContextMenu key brings it back and opens its menu. - A keyboard path for column order and width (R-27, WCAG 2.1.1). With `enableColumnReordering` the column menu offers "Move left" and "Move right"; on a resizable column it offers "Increase width" and "Decrease width" (10 px per step). The menu stays open so a step can be repeated. A move follows the rules of a drop (never across a pinned boundary or out of its header group, never past a grid-owned column); a width is held between the column's bounds and locks the column as the end of a drag does. Each step is announced through the grid's live region; a pointer drag announces nothing. New translation keys, supplied by the host: `grid.move.label`, `grid.move.left`, `grid.move.right`, `grid.layout.increaseWidth`, `grid.layout.decreaseWidth`, `grid.announce.moved`, `grid.announce.resized`, `grid.announce.atMinWidth`, `grid.announce.atMaxWidth`. - A server group row is a row of the keyboard navigation: the arrow keys reach it and Enter on its first cell reaches its toggle. - `Grid.getState()` and `Grid.subscribe(listener)` are part of the published grid: the accepted interaction state, replaced when a change is accepted and handed back unchanged otherwise, and an observer told after every accepted change (ADR-038). - The row and cell shapes are exported from the root: `SourceRow` (a row that stands for a record, with `original: TData`), `GeneratedRow` (a row grouping generated, with `original: undefined`, `groupingColumnId` and `groupingValue`), `SourceCell`, `GeneratedCell` and `GeneratedCellContext` (ADR-037). - The grid options type declares `collationLocale`, the language the `localeText` comparator orders text in. There is no default. - `Column.getNextSortingOrder(multi?)` declares its optional argument: it asks about a toggle that extends a multi-column sort. ### Changed - Persistence writes only the slices it restores (TD-50 d): sorting, filters, visibility, order, sizing, column pinning, pagination, global filter, row pinning and expansion. Row selection, grouping and the resize bookkeeping are no longer written, and remain non-persistent. A stored entry that still carries them loads as before and drops them at the next save. - Cell selection has one model (ADR-043). The active cell is the grid's keyboard position and the anchor of every extension; a range is an explicit selection (a drag that reached a second cell, a Shift extension, a Ctrl/Cmd addition): - A click makes the cell active and stores no range. With cell selection on, the active cell shows the one-cell look. - `onCellSelectionChange` reports explicit ranges only: a click reports nothing, and a single selected cell is `[]`. `aria-selected="true"` is on the cells of an explicit range only. - A drag, Shift+click and Shift+arrow keep the active cell on the anchor and move the far end; Shift+arrow brings the far end into view. They used to move the active cell to the far corner. - Ctrl/Cmd+click adds the cell and keeps the previously active cell selected. Ctrl/Cmd+Shift+click replaces the last range instead of adding an overlapping one (R-52). - Ctrl/Cmd+C copies an explicit range, else the selected rows, else the active cell. - Starting an edit ends the range. - The selection look is drawn from the grid's own state, never from where the browser put the focus, and nothing it draws moves content. The active cell outside a range: one square one-pixel mark at 50% of `--ring`, plus, with cell selection on, the corner cue. A range: an 8% `--ring` fill, a one-pixel outline on its outer edges and one cue on the bottom-right corner of the last range. The active cell inside a range is the window: unfilled, with no mark of its own. When the focus leaves the grid the active cell's look disappears entirely and a range stays; a menu the grid opened changes nothing. The row of the active cell takes no focus tint of its own while a range is on screen. In forced colours the range is Highlight and the active cell a Highlight mark. The cue is a mark, not a handle: there is still no fill and no paste. See `guides/styling.md`. - The cell-range copy (Ctrl+C and the menu) leaves out the grid's own control columns (selection, expander, actions), as the row copy and the exports already did; a range over the selection column no longer copies an empty field for it. - The row menu's consumer builder runs when that row's menu opens, not on every render of every row. - The selection checkbox, the expander and the group toggles are no longer stops in the page's Tab sequence while keyboard navigation is on (R-27): a page of 100 rows used to add 200 stops. Enter or Space on the cell reaches them (action mode), Escape returns to the cell. With `enableKeyboardNavigation={false}` they keep their stop. Action mode now also stops on a cell control with `tabIndex={-1}`; a control a consumer renders in a cell joins this model by declaring that tab index. - The grouping panel is rendered directly above the grid, as the top of the grid's card, instead of between the column headers and the body. The `role="grid"` element owns rows only; the panel's buttons inside it were a critical `aria-required-children` violation. - Controls that exist once per column are named after that column's label (`meta.label`, else a string `header`) instead of its id or a generic word: the filter-row value control (`Filter value for {column}`, with the chosen value when there is one), the funnel and its popover (`Filter {column}`), the clear button, and the column menu button (`Column options for {column}`). The popover's condition select and value fields, the filter builder's selects and fields, and cell editors are named. A multi-column sort names its priority as a phrase instead of a bare digit. Expander and group toggles expose `aria-expanded`, and the group toggle is named by `aria-label` instead of only a `title`. A server group row shows the grouped column's label instead of its field id. - New translation keys, supplied by the host: `grid.filter.valueFor`, `grid.filter.valueForWithValue`, `grid.filter.condition`, `grid.filter.value`, `grid.filter.rangeFrom`, `grid.filter.rangeTo`, `grid.header.optionsFor`, `grid.sort.priority`, `grid.filterBuilder.column`. - R14B complete: the temporary approved-delta test register is retired. Every behaviour it listed is asserted by a directly named test, and the consumer-visible ones are now documented in the API reference: the built-in comparator and predicate names (unknown names, mixed values, `inNumberRange`, `between`, `arrHas`), width bounds, `getRowId`, `postSortRows`, row models and selection scopes, pinned positions, group-row ids and faceting. No runtime behaviour changed. - R14 complete: license and attribution disposition. The package license is unchanged (`UNLICENSED`, proprietary `LICENSE`, `private: true`). The current source contains no vendored or copied TanStack Table material and the package has no TanStack Table dependency, so no TanStack Table notice applies; none was ever present in `LICENSE`. The README License section now states this, keeps the removed fork as history, and notes that the runtime dependency `@workspace/active-virtual` carries its own third-party notice. ADR-007 keeps its historical text with a current disposition. A legal question about earlier distributed versions is escalated (ROADMAP TD-42). Completed temporary rewrite reports were removed; the approved-delta test register is retired in the follow-up R14B. - The development warning for toolbar search on a server grid without a `server.global` column no longer names a removed grid-side resolver. It now says the term is sent to the fetch adapter as `globalFilter` and the server decides which fields it searches. Text only; when it fires is unchanged. - R13 complete: independent technical provenance supported; all 38 canonical documents reviewed, kernel ownership and API guidance corrected, neutral examples enforced by a scoped documentation guard. Historical ADR evidence and approved deltas remain through R14. Five internal hook aliases now say KernelGrid; source comment corrections do not change runtime behavior. Attribution and license decisions remain with R14. Performance follow-up remains due 2026-09-29; unrelated repository validation debt is recorded in the R13 closure report. - The grid runs on its own kernel; the replaced engine is no longer used at runtime (ADR-038). The published objects, options and state slices keep their names and shapes. - A callback that is handed a row's record — `meta.valueFormatter`, `meta.editable`, `meta.cellClassName` as a function, `meta.cellClassRules`, `meta.cellStyle`, `meta.tooltipValueGetter`, `meta.tooltipField`, `meta.colSpan`, `onRowClick`, `onRowDoubleClick`, `onCellClick` and the row `contextMenu` — is no longer called for a row grouping generated, which has no record; it was handed `undefined`. On such a row no edit starts and no row menu opens. Group totals, group footers, server-grouping aggregates and the presentation export of an aggregate are shown through the column's declared semantics (`semanticType`) or its `aggregatedCell` template and no longer pass through `valueFormatter`, which received an empty object in place of a record. Ratified as delta DC-1 (2026-09-22): a column whose aggregate relied on `valueFormatter` for its notation declares `semanticType` instead. - The `median` aggregate skips an absent value (`null`, `undefined`, `''`, `NaN`, `±Infinity`) and is the median of the values that are present; without a present value it has no result. It used to return nothing as soon as one value was absent. Ratified as delta DC-2 (2026-09-22). - **Group-row aggregation is declarative.** A column shows an aggregate in a group row only when it declares `aggregationFn`. A numeric column without a declaration used to be summed through an implicit `auto`; it now shows an empty cell, because a numeric value (an identifier, a year, a rate) does not imply an additive meaning. A column that should total declares it, for example `aggregationFn: 'sum'`; `auto` stays available as an explicit choice. Ratified as delta DC-3 (2026-09-22). - A mounted client grid no longer builds a row object for every filtered row on each data change. The pagination total and the status announcement read the filtered count without building rows, and the row numbering is built only when a row-number column is shown. At 100 000 rows this removes 200 000 of the 200 010 row objects a data change built. - `onSelectionChange` hands over the records of the selected source rows only. - **Breaking (types).** `Row` is `SourceRow | GeneratedRow`, and `row.getIsGrouped()` narrows it. Code that reads `row.original` where a generated row can arrive — a comparator, an eligibility predicate such as `enableRowSelection`, `renderDetailPanel`, `getRowCanExpand`, a row model — narrows first: `(row) => !row.getIsGrouped() && row.original.active`. Callbacks that only meet rows standing for records keep `original: TData`: a filter function receives a `SourceRow`, an aggregation's `leafRows` are `SourceRow`s, the parent `getRowId` receives is a `SourceRow`, and a body-cell renderer's `CellContext` carries a `SourceRow`. An `aggregatedCell` renderer receives a `GeneratedCellContext`. `Cell` is `SourceCell | GeneratedCell`. - A persisted pagination state whose page size is not a whole number greater than zero is discarded on restore, as a zero or negative size already was; a fractional size is no longer restored. One rule decides every page size. - `createLocaleSortFn(locale)` refuses an empty or missing locale with a `TypeError` instead of falling back to the host locale. The sorting semantics for a given locale are unchanged. Its declared return type is now a structural comparator over `row.getValue(columnId)`; the parameters are unchanged. - `createColumnHelper` and `createLocaleSortFn` are implemented independently of the engine; the objects the helper returns are unchanged. - The objects a consumer receives (`Grid`, `Row`, `Column`, `Cell`, `Header`, `HeaderGroup`, `CellContext`, `HeaderContext`), the state slices, the callback types and `DeepKeys` are declared by the package itself instead of aliasing the engine's types (ADR-036). Every supported member keeps its name; the row, cell and callback signatures that change are listed above (ADR-037). Collections the grid hands out (`getRowModel().rows`, `getAllColumns()`, `row.getVisibleCells()`, …) are typed `readonly`; the value type of a column now reaches `Column`, `Cell`, `Header` and both renderer contexts, and every slice reset declares its optional `defaultState` flag. - `createLocaleSortFn` and `createColumnHelper` are documented as part of the root export; they were described as contributor-only. ### Fixed - **A replaced column definition is what the grid shows.** Re-rendering a mounted grid with the same records and a replaced `accessorFn`, `accessorKey`, `cell`, `header`, `aggregatedCell`, `meta` (for example `valueFormatter`, `semanticType`, `cellClassName`) or default column now updates the visible cells and headers when the render commits. The grid already answered with the new definition; the cells went on drawing the old one until an unrelated interaction. No row is rebuilt for it: the displayed rows keep their identity, and only their cells are rebuilt, also in a virtualized body. - **The active cell is marked as soon as a click makes it active**, not only after an arrow key moves it (styling in `@workspace/active-ui`). - **A sort label clicked with a held Shift (multi-sort) no longer shows the keyboard focus ring.** Keyboard focus on a sort label still shows it. - A malformed page size in the address (`pageSize=2.5`, `25junk`, `1e2`) is ignored instead of being truncated into a different, valid size; the one page-size rule reads the text (PW-02). - A malformed page number in the address (`page=2junk`, `1.5`, `-1`, `0`) is ignored instead of being truncated or clamped into a page; it follows the same textual rule as the page size. - With two or more sorted columns every sorted header shows its priority, the first one included; the `1` was never shown, so a multi-sort read as `2` without a first. With one sorted column no number shows. The header reads the priority from `Column.getSortIndex()`. - Clearing a text filter whose operator is not the column default (for example "does not equal") clears it; the debounced input used to re-apply the old text under the default operator. - Server grouping no longer hands its group headers and load-more markers to `getRowId`; the grid names them itself. - Server grouping shows the group rows the server returns. Since the kernel cutover the grid grouped those rows a second time by the mirrored grouping and folded every group into one empty group; in server-grouping mode the grid's own grouping is now manual. - A page restored from the address or from the stored state survives the mount on a grid with a filter row. Every filter control clears its own column once on mount, and that no-op write returned the grid to page 1; a sort, filter or search now returns to the first page only when it changes the slice. - The grid type declares `setColumnResizing`, the member the grid has at runtime, instead of the misspelled `setcolumnResizing`. ### Removed - The CSS token `--grid-z-cell-focus`. The active cell's mark is an overlay inside the cell and needs no stacking of its own; the old rule's z-index also lifted a scrolled active cell over the sticky pinned columns. - The active cell's inset box-shadow ring on `.grid-cell[role='gridcell']:focus`, replaced by the overlay mark above. A consumer rule that restyled that shadow no longer has an effect. - The translation keys `grid.copy.label` and `grid.copy.withHeaders`: the one ambiguous "Copy" is replaced by the items named by scope (see Added). - The translation key `grid.filter.openFilter`: every funnel is named with its column (`grid.filter.openFilterFor`). - From the public types: members that belong to the engine's implementation (`_features`, the prototype and row-model fields, value caches, `autoResetExpanded()`), the state-container fields `baseStore` and `store` (grid and options), the `debug*` options and `mergeOptions`, and five declared grid members that do not exist at runtime (`getFacetedRowModel`, `getFacetedUniqueValues`, `getFacetedMinMaxValues`, `getPinnedLeafColumns`, `getPinnedVisibleLeafColumns`). `Header.rowSpan`, which is always `0`, is no longer declared. - The dependency on `@workspace/active-state`. It served only the replaced engine, which is removed from the source together with its binding (TD-42, R12); neither was reachable from the published entry, so no import changes. ### Documentation - The version-to-version migration guides (`migration/from-pre-1.0.md`, `v1-to-v2.md`, `v2-to-v3.md`, `v3-to-v4.md`) are removed: they described upgrades to versions and helpers that no longer exist. A current integration starts from the current API; this changelog keeps the history. The AG Grid migration guide (`migration/from-aggrid.md`) is removed too: the package is private and has no external adopters. - `guides/i18n-date-filters.md` is merged into `guides/filters.md` and `DATA-ACCESS.md` into `ARCHITECTURE.md` § Datatoegang en eigendom. - The README quick starts now pass the required `locale`, and `api/filters.md` names the real operator label keys (`grid.operators.*`). --- ## \[10.2.2\] - 2026-09-21 - The baseline is frozen on a quiet machine, and the column window is bounded > **R-26 is closed.** The entries below were held at 10.2.1 until a baseline > could be frozen on a quiet machine: freezing the numbers of a loaded machine, > or of a defect, is worse than having no baseline at all. On 2026-09-21 both > rounds reported `quiescent: yes` -- the browser round with only admin:3002 > and auth:3000 listening, the jsdom round with nothing listening at all -- and > `benchmarks/baseline.json` now exists at this version. The immediate > compare-back was green: no counter drift, no scenario-integrity drift. ### The benchmark baseline exists (R-26) `benchmarks/baseline.json` is the package's first legitimate frozen baseline: both runners, 18 scenarios, 3 warm-ups plus N = 15, raw observations retained, every counter observed fifteen times and equal among itself, and the machine condition of each round recorded inside the document alongside an explicit human confirmation. #### Added - **`benchmarks/baseline.json`** -- the frozen reference point. It is the one shared artefact under `benchmarks/`; everything under `benchmarks/local/` stays ignored. #### Measured - **Zero counter drift against the 2026-08-28 trial rounds**, in both runners. Those rounds were measured on a machine carrying seven to nine dev servers with coefficients of variation up to 89%, and their counters are still exactly these. That is the tightest evidence so far that the counter class is genuinely machine-independent while the timing class is not. - **Memory is a usable trend for the first time.** The jsdom round carries `memoryPolicy: "trend"` and all twelve retained-heap series sit below a 12% coefficient of variation, where 10.2.1 published negative deltas of hundreds of megabytes. It remains a trend and does not become a gate. - **The three pre-rewrite measurements R1 asked for: one delivered, two reported as unobtainable.** Accessor-call counts and engine-wide publication counts do not exist in the shipping engine, and adding a scenario counter is by the scenario manifest a baseline change -- which would have invalidated the very reference point this work produced. The column-normalization question is answered as a static census instead: of 240 `columns={...}` bindings across 235 consumer files, 40 rebuild the array identity on every render and therefore miss the engine's column memo every render. Full reporting in `PERFORMANCE.md`. ### The public column contract is a contract again (TD-31B, ADR-035) `GridColumnDef` resolved literally to the top type, so `const columns: GridColumnDef[] = [42, 'not a column']` compiled. There were two independent causes with that effect, and repairing one of them did nothing for the other. No runtime behaviour of this package changed. #### Changed - **The feature set is supplied, not hidden.** `ReactLayerFeatures` was `any`; it is now `StockFeatures`, in one named package-private type. Measured: `CoreFeatures & StockFeatures` produces 16 errors that `StockFeatures` does not and removes none, all of them an alias claiming a wider set than the engine builds. - **`GridColumnDef` is a union of three kinds** — accessor-key, accessor-function, and group-or-display. A display column is not a fourth kind on purpose: as a separate member it would absorb a group whose children are invalid. - **`accessorKey` must name a field of the row**, dotted paths included, computed from the unwidened row type. - **A group's `columns` is recursively this same definition**, so a nested child carries `type`, `server` and `meta` exactly as its parent does. It was the engine's own shape and carried none of them. - **`createColumnHelper` declares the public shapes** and infers each column's value type from the key or the function. It is the only construction path that infers a value type at all. - **`enableRowSelection`'s predicate receives the ROW, and the prop now says so.** It was declared with the record while the engine passed the row, so a predicate written against record fields read `undefined` and refused every row. This is a corrected declaration, not a behaviour change in the package; the one consumer using the function form was corrected with it. - **`meta.statusTones` is declared partial.** A total `Record` claims every possible string has a tone, which contradicts the documented fallback, and it refused a shape TypeScript synthesizes when two columns with different key sets share one array literal. - **The header selection checkbox states its value honestly.** It passed `'indeterminate'` through a boolean field, so a partial page has always rendered as fully checked; the rendering is unchanged and the missing third state is `ROADMAP.md` TD-45. #### Added - **`hasAccessorKey` / `hasAccessorFn`** — the two narrowings for reading a definition back. `accessorKey` is not reachable on a union, and a local `column as {accessorKey?: string}` keeps compiling after the field moves. #### Removed - **`BivariantGridColumnDef`** — the bivariance it provided is now the default of the public type itself, so a second name for the same type would be a second truth. Its five consumers name `GridColumnDef`. - Three `(r: any)` annotations in component code that existed only because the React layer hid `TFeatures`. --- ### Source language convergence completed, and the engine's provenance measured (AGRID-STD-04) No runtime behaviour changed and no public API changed. Every change is authored prose: comments, docblocks, test descriptions, `it.each` labels, assertion messages and developer diagnostics. #### Changed - **All remaining Dutch authored prose is English (TD-44, closed).** Three independently constructed sweeps found 908 lines across production source, tests, `e2e/`, `benchmarks/` and `scripts/` -- more than double the 242 the previous re-measurement reported, because a comment-only scan cannot by construction see `describe`/`it`/`it.each` descriptions, assertion messages or Dutch `throw new Error(...)` prose. Seven English comments that quote a Dutch UI label remain, as `CODING_STANDARDS` § 2 permits, along with the Dutch UI expectations and locale fixtures themselves. - **Benchmark scenario descriptions are English.** `scenario-manifest.ts` carried Dutch `measures` prose that lands in every published benchmark document. The harness assertion that matched on that Dutch text moved with it, so the guard keeps its subject. #### Documentation - **TD-42 is reframed from attribution to provenance.** Measured: `src/engine/**` (100 files, 14,004 lines) and `src/react/**` (4 files, 392 lines) are a fork with a rename of TanStack Table v9 `table-core`, not an independent implementation -- established on directory layout, per-feature file names, the complete identifier set, and two character-for-character identical helpers in `utils.ts`. 27% of the package is derived; the remaining 73% is independently authored active-grid code. `LICENSE` was not touched, no attribution text was added, and no existing provenance reference was removed. The evidence and the seven-tranche replacement plan are in `docs/research/ACTIVE-GRID-TANSTACK-PROVENANCE-ASSESSMENT-2026-09-20.md` and in `ROADMAP.md` TD-42. This is a technical finding, not legal advice. ### Documentation drift corrected (AGRID-DOC-01) No runtime behaviour changed and no public API changed. Listed here because three of these corrections were misleading to a consumer, which is the one case the changelog convention records a documentation-only change for. #### Documentation - **The README no longer says the package is MIT.** `package.json` declares `UNLICENSED`, `LICENSE` carries the proprietary Active Development text and `private: true` is set. The README also claimed the TanStack attribution was "preserved in `LICENSE`"; it is not there, and no file under `src/engine/` carries an upstream copyright header. ADR-007 requires that clause, so the gap is now stated in the README and registered as `ROADMAP.md` TD-42. - **`next` is no longer described as optional.** `ARCHITECTURE.md` said "optioneel Next.js 16" and `FUNCTIONEEL-OVERZICHT.md` said "optioneel Next.js 15 of 16", both left over from before AGRID-F04 made it a required `^16.0.0` peer. - **`docs/api/columns.md` typed a column group's `columns` as `GridColumnDef[]`.** It is the engine shape, `Array>`, and has never been otherwise: a group child carries no `type`, `server` or `meta`, and its `TValue` is pinned to `unknown`. Repairing the type was TD-31B, which landed in this same release and made the field the public definition at every depth; see the TD-31B entry above for what it is now. - **`docs/guides/filters.md` described `withEmpty` against a symbol that does not exist.** `toPrismaWhereFromContract` has no occurrence anywhere in the repository, and the claimed "silently dropped with a dev-warning" does not happen: in `@workspace/active-resources` a non-nullable text column matches the empty string, and number, date and boolean have `isEmpty` rejected by the operator allowlist before any condition is built. The same wrong sentence is corrected in the `operator-presets.ts` docblock. - **`docs/FUNCTIONEEL-OVERZICHT.md` rewritten against the measured package.** It still described a different monorepo, counted 7 e2e specs where there are 17, cited the withdrawn ADR-011 guard, named a `--max-warnings 545` cap that is 100, and listed the Prisma query builder and the schema-aware server contract — both removed in 7.0.0 — as current differentiators. Batch editing, undo/redo, PDF export, the filter builder and cell/range selection were missing and are now listed. - **Version labels.** ADR-030, ADR-031 and ADR-032 read "geaccepteerd (…, 10.2.2)" as though 10.2.2 existed. Nine places now say what `INDEX.md` already said: the work sits in `[Unreleased]` and ships as 10.2.2 with the R-26 freeze. #### Removed - **Eighteen temporary documents**, after each was read in full and its durable content handed to its owner: `docs/plans/`, `docs/traycer/`, `docs/comparison/`, `docs/integrations/`, six research files, the resize handle analysis and the 2026-07-21 audit verification — about 12 000 lines. active-grid no longer has a work-archive or history concept; git is the history. Two research files remain, each with a deletion trigger recorded in `docs/INDEX.md`. ### The package conforms to `/docs/standards` (AGRID-STD-03) No runtime behaviour changed, no public API changed, and `lint:api-drift` reports the same 271 exports and 401 documented symbols as before. The full suite runs green on the same 193 files and 2 808 tests. #### Changed - **Braces on every conditional and loop body** (CODING\_STANDARDS § 5), across all authored source: 825 braceless bodies in 182 files became zero. Measured over production, engine, test, benchmark, e2e and script source -- § 1 puts all of them in scope -- and applied as an AST rewrite with `eslint --fix` running `curly` alone, followed by Prettier. `src/engine/` is included: ADR-007 vendors it and says to treat it as internal code, no file under it carries an upstream copyright header, and there is no generator or synchronisation contract that § 2 would exempt. The TanStack portion clause in `LICENSE` is untouched. - **All sixteen remaining hooks declare an explicit return type** (CODING\_STANDARDS § 6), and **not one name was added to `src/index.ts`**. Four reuse a type that already existed; the other twelve get a result interface in their own module, fourteen of which are module-private. TD-39 argued that naming a hook result inevitably publishes it -- the package already contradicts that six times (`UseAutoSizeReturn`, `UseEditHistoryReturn`, `UseServerGroupingReturn`, `UseServerSelectAllReturn`, `UseBatchEditingStateReturn`, `UseGridImperativeResult`), none of which appears in the barrel. Exported non-component functions in the non-engine source now carry an explicit return type 205 out of 205 times. - **`curly` is enabled in the package's own `eslint.config.mjs`** so a regression shows up in `pnpm lint`. It reports as a warning, because the shared base downgrades every severity; blocking enforcement is a repository decision and is registered as TD-40. #### Documentation - **Authored source prose is English** (CODING\_STANDARDS § 2): 1 417 comment and docblock units across 185 files, 154 `describe`/`it`/`test` descriptions, and two developer-facing `throw` messages. Meaning-preserving translation only -- no claim added, no commentary rewritten, and every date, identifier, path and translation key left exactly as it stood. The Dutch UI labels quoted as expectations in `column-chooser.spec.ts`, `column-resize.spec.ts` and `context-menu.spec.ts` stay: those specs run against `apps/admin`, which hands the grid its own `t`, and § 2 covers test expectations whose language is the point. #### Removed - **`bash.exe.stackdump`** (1 196 bytes, an msys bash crash trace). Already git-ignored, referenced by no gate, script or document; its call stack was read before removal and contains no Datyx frame. ### A fallback message speaks the current language, and the manifest tells the truth (AGRID-STD-02) #### Fixed - **A translated fallback no longer arrives in the language the user just left (AGRID-F03).** Three `useCallback`s reached for the grid's resolved translator without listing it: the paged fetch and the infinite-scroll fetch in `useDataOrchestrator`, and the cell flush in `useBatchEditingState`. A Datyx host memoizes `t` per language, so the identity rotates exactly on a language switch and the callbacks kept the previous one. Reach: `grid.error.fetchFailed` (a fetch that rejects with a non-`Error`) and `grid.edit.validationFailed` (a column validator that throws). `translate` is now in all three dependency arrays; the fetch trigger already calls through `fetchDataRef`, so nothing refires. #### Changed - **`next` is a REQUIRED peer dependency (AGRID-F04).** `peerDependenciesMeta` marked it optional while `LinkCell` imports `next/link` at module scope and the single entry point exports `LinkCell` -- so importing the package has always had to resolve `next`, and `@workspace/active-ui` requires it too. The manifest and the README now say what the package actually needs. No export, subpath or component moved; no consumer changes. - **Type-only imports use `import type`** in 30 modules (CODING\_STANDARDS § 4). Mechanical; no runtime or type-contract change. - **Five production `as any` casts and their suppressions are gone from `export-utils.ts`.** `jspdf-autotable` v5 publishes `UserOptions`, `RowInput` and `CellHookData`; the prototype-installed `autoTable` is now reached through one named view type instead of `any`, and the two `didParseCell` hooks are typed. The PDF option objects and both table bodies are checked against the package's own types for the first time. - **Thirteen public functions declare an explicit return type** (CODING\_STANDARDS § 6): the twelve `exportTo*` variants and `copyToClipboard`. The sixteen that remain are all hooks whose result shape would have to be NAMED, which is a public-API decision; registered as TD-39. - **Ten `react-hooks/exhaustive-deps` suppressions carry an English reason** (CODING\_STANDARDS § 10). Each was probed by removal first: all ten are still necessary. Two redundant `useMemo` dependencies that the rule flagged as unnecessary were removed from `grid.tsx` and `use-grid-setup.ts`; neither is read in its memo body. active-grid now lints at zero warnings. #### Documentation - **TD-37: two stale source comments corrected.** `types/filters.ts` called the module-scope operator lists "NL defaults" -- they carry the package's own standalone English table. `i18n/grid-translation-context.tsx` and `types/props.ts` both claimed the built-in calendar mapping "covers en/nl/de/fr/es"; since ADR-034 `utils/date-locale.ts` defers to active-i18n and active-grid carries no language list at all. ### Keyboard navigation and auto-height read the live grid model (AGRID-CORR-01) #### Fixed - **Keyboard navigation no longer freezes its bounds on the first render.** `useKeyboardNavigation` derived `rowCount`, `colCount` and `pinnedTopCount` in a `useMemo` keyed on the grid instance, which never changes identity (`react/useGrid.ts` builds it in a `useState` initializer). The counts therefore stayed on whatever the first render saw. Consequences, all gone: - **`mode="server"`**: the first render has zero rows by contract, so `rowCount` was 0, `Math.min(rowIndex + 1, rowCount - 1)` targeted row -1, and the first ArrowDown after the fetch threw DOM focus out of the grid. Every later key then landed nowhere. WCAG 2.1.1 is level A. - **`mode="client"`**: a row model that changed length after mount — filtering, searching, a shorter last page, data arriving late — kept the old bound, so arrows, `End`, `Ctrl+End` and `PageDown` could not reach the rows that were there. - **The column axis**: the column chooser, responsive hiding and a restored visibility state changed the visible column count without changing the bound, so `ArrowRight`/`End` stopped short or aimed at a column that had been hidden. - **`pinnedTopCount`**: pinning a row to the top while the grid was mounted left `scrollToIndex` converting to the wrong center index. - **`meta.autoHeight` now takes effect when its column becomes visible.** Both bodies (`sections/body.tsx`, `sections/body-virtual.tsx`) derived `hasAutoHeight` from the same frozen memo. A column with `autoHeight` that was hidden at mount — a persisted visibility state, responsive hiding — and made visible afterwards left every row on a fixed height, clipping content that should have grown. No public API change, no remount, no new state, no change to `@workspace/active-virtual`. The counts are read during the render and the auto-height flag hangs on the visible-column model; both are engine-memoised on the state that can change them, so an unchanged model costs a dependency comparison. ADR-030 carried the rule already and now names `grid` as the second stable instance. ### The date calendars mark the presentation's day, not the runtime's (AUI-FINAL-REM-01) #### Changed (breaking) - The calendars of `DateEditor` and the date filter pass `today` to active-ui's `Calendar`, which now requires it (active-ui ADR-025, amendment of 2026-09-19). The day comes from the zone of the grid presentation, through the same `todayInPresentation` the relative date presets use; without a selection the calendar opens on that day's month. `new Date()` is gone from both. - A grid presentation **without a time zone** (a bare language code) now throws a configuration error when a date editor or a date filter calendar renders, the rule the relative presets already followed. Every product grid is fed by `usePresentation()` under a resolved zone and is unaffected. ### The host owns words and presentation; no invented English (I18N-HC-02, ADR-034) #### Changed (breaking) - `locale` on `ActiveGrid` and `LookupGrid` is **required**. `useGridLocale()` has no `'en'` default any more and throws a configuration error without a presentation. - A host translator (`t`) is **authoritative**. A key it does not carry renders as the raw key; the former silent replacement by the English table (`result === key`) is gone. - The fetch, save, validation and cell-error fallbacks ask the same translator (they called the English table directly before). #### Fixed - Ctrl+C of selected rows or of the focused cell copied values in English notation whatever the grid's `locale` (the cell-range copy was correct): `useKeyboardNavigation` read the presentation context from outside the provider the grid renders. The grid now passes its `locale` to the hook. ### The scrollbar is the sidebar's, pixel for pixel (UI-SCROLLBAR-02, active-ui ADR-026 amendment) #### Changed - The scrollbar lane is built exactly like the house `ScrollArea` bar: `padding: 1px`, a transparent 1px border on the side facing the content (`border-left` vertical, `border-top` horizontal) and a `flex: 1` thumb, replacing the symmetric 1.5px inset (`--grid-scrollbar-inset` is gone). The thumb therefore has the sidebar's width and position (7.36px at DPR 1.5625, 1px from the outer edge) instead of a 7px thumb centred in the lane. Measured side by side in UI Lab `scroll-area#case-grid-parity`: lane, thumb, inset, corner, rest, hover and scrolling state identical, light and dark. - No change to the adapter, the lane width, the colours, virtualization or the scroll math. `scrollbar-presentation.test.ts` pins the construction. ### The scrollbar reads the platform contract (UI-SCROLLBAR-01, active-ui ADR-026) #### Changed - `--grid-scrollbar-thumb`, `--grid-scrollbar-thumb-hover` and `--grid-scrollbar-track` default to the platform tokens `--scrollbar-thumb`, `--scrollbar-thumb-hover` and `--scrollbar-track` instead of their own values. The resting colour is the same (`--border`). - **The thumb now reacts under the pointer and while scrolling**, with the platform hover token. AG-SCROLL-01 kept it at rest only because the house `ScrollArea` did not react either; both now do. `theme.scrollbarThumbHover` still overrides it per grid. - Under `forced-colors` the thumb is `CanvasText`; before, the browser replaced its background with the lane colour and it disappeared. - No change to the adapter, the lane, the geometry or virtualization. `scrollbar-presentation.test.ts` pins the aliases and the hover rule. ### Grid fields carry the platform field focus (UI-FOCUS-01, active-ui ADR-022) #### Changed - **Every grid field focuses like a form field**: a border at 50% of `--ring`, no ring, in light and dark. The popover filter inputs and selects (text, number, interval, date, operator) dropped `focus:border-primary focus:ring-2 focus:ring-primary/20`; the date popover's text inputs dropped the same with `ring-1`; the filter-builder fields dropped `focus-visible:ring-1 focus-visible:ring-ring`. All of them take the internal `FIELD_FOCUS_CLASS` (`components/primitives/field-focus.ts`), which is bound to active-ui's `Input`. - **The inline filter bar is dimmed** through active-ui's `--grid-filter-trigger-focus-border`, now `color-mix(in oklab, var(--ring) 50%, transparent)` (was `var(--ring)`). The dead `focus:border-primary/50 focus:bg-background` on the inline text and number inputs is gone; the grid CSS already won over them. - The toolbar search no longer adds `focus-visible:ring-1` to `Input`, and the page-size `SelectTrigger` no longer carries a no-op `focus:ring-0`. No public API change. Buttons, cells and the grid container keep their focus. Guarded by `field-focus-contract.test.ts`. ### `±Infinity` is absent, like `NaN` (active-i18n F3b-9) #### Changed - **A non-finite number is absent on every layer.** `isAbsentCellValue` now treats `Infinity` and `-Infinity` like `NaN`: the cell draws the placeholder instead of `∞`, `isEmpty` matches, the presentation export writes `''`, and `sum`/`mean`/`min`/`max` skip the value (`median` returns no result). The DATA export was already empty. Walt's decision A2: the human formatters of active-i18n return `''` for a non-finite value, and no product means an infinite number. Supersedes the REVIEW-AG-004 rule that Infinity is data. The internal `toNonFiniteNumber` pass-through in the presentation export is removed. ### Date filter: a calendar day is not an instant (active-i18n F3b-3) #### Fixed - **A calendar-day column no longer filters one day early west of UTC.** The client-mode date filter read every cell as an instant in the presentation's zone, so a Prisma `@db.Date` or `…T00:00:00.000Z` value for 9 March matched 8 March on an `America/Curacao` surface. A `semanticType: 'date'` column (and a `filterType: 'date'` column without semantics, as `DateCell` shows it) now compares calendar days without a zone. ADR-033. - **The filter builder receives the presentation and each field's date semantics**, which it did not before. #### Changed - **A `dateTime` column filter and the relative date presets require a time zone in the grid presentation** and throw a `TypeError` without one, instead of reading the browser's day. `DateFilterPreset.getValue` documents this. - `evaluateDateFilter`, `evaluateFilter`, `advancedFilterFn` and `evaluateFilterExpression` take an optional date-semantics argument. ### `SelectFilter`: a declared selected state and reader-language search (`DATYX-FINAL-013`) #### Added - **Every option row in the select filter declares its membership.** An option is a `