`aria-*` boolean attributes serialize as the empty string

Lisez design/v2/follow-up-issues.md et inspectez le chemin de sérialisation de Base pour les attributs aria sur la branche 2.0. Exécutez la golden HTML suite et examinez les snapshots concernés du catalogue et des composants, notamment les séparateurs de breadcrumb, les points de suspension, CommandInput et les SVG décoratifs. C’est terminé lorsque les booléens aria sont rendus sous la forme "true" ou "false", comme décidé, que les snapshots 2.0 ont été réenregistrés et que leur diff a été examiné.

• Where: the components that pass a boolean under the aria hash:
  aria: {hidden: true} in BreadcrumbSeparator, BreadcrumbEllipsis and
  PaginationEllipsis (also reached through DataTablePagination),
  aria: {disabled: true} in BreadcrumbPage, aria_expanded: true in
  CommandInput. The breadcrumb, pagination, data_table and command
  snapshots carry aria-hidden="", aria-disabled="" or aria-expanded="".
  Components that pass the String "true" — InputOtpSlot,
  NativeSelectIcon, AlertDialogContent — already serialize correctly and
  are not affected.
• Effect: the ARIA value grammar accepts true/false/undefined; an
  empty string is invalid and browsers resolve it as not set. The breadcrumb
  and pagination separators and ellipses are therefore exposed to assistive
  technology, "current" breadcrumb items are not announced as disabled, and the
  command input's expanded state is unset. This is Phlex's serialization of
  true under a nested hash (a bare attribute), so any component that passes
  a boolean there is affected — five today.
• Fix: pass "true" (a String) for aria-* attributes, or add a
  serialization rule in Base that stringifies booleans under the aria
  key; re-record (on the 2.0 line) everything that changes and review the diff.
• 2.0 note: spec §4.3 currently preserves this behaviour for parity
  (true → ""). The right 2.0 behaviour is aria-x="true"; decide it
  explicitly and let the re-record (on the 2.0 line) show the blast radius.

Surfaced by the golden HTML suite (PR #536), which pins what 1.6 renders today — this defect included. Per maintainer decision (2026-09-19) main stays as is: this is addressed on the 2.0 line, and the golden snapshot that pins it is re-recorded there with the diff reviewed.

Source: design/v2/follow-up-issues.md.