Accessibility

Calling something accessible is easy and usually wrong, so here is the split.

Built in

• <opengrid-table> and <opengrid-pivot> are native <table> elements. <opengrid-grid> uses role="grid", because it takes over the arrow keys — for a report table that would be the wrong promise, which is why the table exists.
• Every feature is keyboard-operable. The grid follows the WAI-ARIA grid pattern; the key matrix is in The public API.
• Every state change goes through one polite live region.
• The component’s own words are English and overridable with set_texts. A lang given with them goes onto the elements that carry those texts, never onto the data.
• A pivot group whose value is NULL is named (no value), one whose value is the empty string (empty): an empty header cell is silence to a screen reader.

Verified by tests, on every commit

Roles, accessible names, aria-rowcount / aria-rowindex / aria-sort / aria-selected, the roving tabindex and the keyboard matrix, target sizes, and the language of the component’s words versus the page’s data. The Playwright suite in tests/e2e/ runs axe-core over the rendered elements.

tests/e2e/announcements.spec.js records the status line’s successive states, so an announcement that fires twice, never, or too early to survive the next result is a failing test. The first time it ran, it found that turning a page announced nothing: the row count is the same on every page, so the live region repeated itself while the content changed.

Not yet verified

A screen reader has not been through the finished V1. The protocol exists and is the last open item before a release. CI runs the e2e suite in Chromium; before a release, just e2e-browsers runs it in Firefox and WebKit too, at a desktop viewport.

So: the structure is tested hard, the experience is not signed off. Until it is, treat the accessibility of this library as well-built and unaudited.