[pournasserian]

Closes #2582

What

The fourth slice of F-DSK (#2422): keyboard-first batch review (DSK-04, UX-02) on the merged interop module (#2500). It also adds the Playwright browser tests and the automated WCAG 2.1 AA checks (UX-04) that #2434 left for this epic.

• admin.js roving lists: a [data-ui="rove"] list of [data-ui="rove-item"] elements uses the existing rove helper.
  • J and K move to the next and previous item from anywhere outside a field, a menu or another modal. From outside the list, J enters at the item in the tab order and K at the last one.
  • The arrow keys, Home and End move while focus is in the list.
  • The list is read on every key, so items DiffView adds in a later batch are reached too.
• Desk:
  • The rows rove, with only the first in the tab order, and Enter opens a set.
  • Keys 1 to 6 switch queues through aria-keyshortcuts.
  • A set link carries its queue (sets/{id}?queue=awaiting-me).
  • When the set view comes back to an empty queue, the desk says what was done.
• Set view:
  • The diff items rove: DiffView gets an opt-in Roving parameter, used by the island. /ui and other uses keep every summary in the tab order.
  • Shortcuts: A approves, R opens the reject dialog, P publishes, N opens the next set in the queue and B goes back. The keys show on the buttons and are hidden from screen readers, which read aria-keyshortcuts.
  • After an approve or reject from a queue, the page redirects to the next set in that queue with "The previous set was approved", or back to the queue once it is empty.
• ShortcutHelp (Components/Review): a dialog opened with ? that lists the page's keys with Kbd, on the desk and the set view.
• TableBody passes extra attributes through, for data-ui="rove".
• Accessibility fixes from the axe run on /ui:
  • CommandSeparator is aria-hidden instead of role="separator", which a listbox may not hold.
  • Kbd text uses text-foreground, because text-muted-foreground on bg-muted was 4.3:1 in the light theme.
• CI: caches ~/.cache/ms-playwright, installs Chromium with playwright.ps1 install --with-deps chromium after the build, and sets FLUENTCMS_TEST_BROWSER. CONTRIBUTING says how to run the browser tests locally.

New test packages

• Microsoft.Playwright 1.63.0 (MIT), in FluentCMS.IntegrationTests only.
• Deque.AxeCore.Playwright 4.14.0 (MIT), which runs axe-core in the page, in FluentCMS.IntegrationTests only. It brings Deque.AxeCore.Commons (MPL-2.0), which carries axe-core itself.

Neither ships with the product. MPL-2.0 is file-level copyleft and applies only if the package's own files are changed and distributed, which this repository does not do.

Why

• Browser tests in the integration project, behind a variable. They need the real host, AdminHost and the same data setup as the screen tests. Skipping without FLUENTCMS_TEST_BROWSER matches the PostgreSQL tests, so dotnet test still needs only the .NET SDK.
• J and K as well as arrows. Arrows only move inside a list, so they never take over page scrolling. J and K reach the list from anywhere with one hand, which is what makes a ten-item review fast.
• Move on after a decision. A batch is a queue of sets. Going straight to the next set saves a round trip per set, and the decision is already saved before the redirect.
• The island says when it is interactive. When the circuit attaches, Blazor replaces the prerendered items, which drops focus from a summary that J reached before then. ChangeSetItems marks itself data-interactive="true", and the keyboard test waits for it. A full run under heavy local load lost focus that way before the marker was added. On a reviewer's screen this window is the circuit's start, usually well under a second.
• Not in this PR: per-item approval of a routing sample (RTE-04) needs routing's sampling, so this slice approves whole sets. A revert shortcut is left out because a single key should not open a revert by accident.

Tests

Browser (Playwright, headless Chromium, against the host):

• UX-02: with the keyboard alone, the reviewer signs in, presses J, opens the ten-item set and steps through all ten items. Only the focused summary is in the tab order, and K moves back. A approves the set and lands on the next set in the queue with the done note, in well under two minutes (asserted), and the set is approved.
• Reject by keyboard: 2 switches to awaiting anyone. J and Enter open the set. R puts focus in the reason, and typing, Tab, Tab and Enter reject it. The reviewer lands on the now empty queue with the note, and the set is rejected.
• Typing and the help: letters typed in the reason field are not shortcuts. ? opens the help, and A does nothing while it is open. B goes back to the queue.
• WCAG 2.1 AA: axe-core (wcag2a, wcag2aa, wcag21a, wcag21aa) reports no violation on /ui, the desk, the stale facts queue and the set view, in LTR and RTL, light and dark.
• A smoke test that admin.js is served by the test host.

bUnit:

• DiffView with Roving marks the list and summaries, and only the first summary is tabbable. Without it, nothing changes.
• Desk rows rove, queues have keys 1 to 6, the help lists five keys, set links carry the queue, and the done notes.
• Set view: A and R are declared, back and next links follow the queue with B and N, the done note shows, and the help lists eight keys. Without a queue, it goes back to the desk and has no next link.

Integration: approving from a queue redirects to the next set with done=approved, and rejecting the last one redirects to the queue with done=rejected, which the desk shows.

dotnet build -warnaserror, the full suite with the browser tests, and markdownlint pass locally against a postgres:17 container as well as SQLite. One loaded full run had EntityRelationsTests (PostgreSQL) fail once; it passed three reruns, and this PR does not touch it.

Checklist

• Linked to an issue a maintainer has confirmed
• Targets v2
• Tests or docs updated where behaviour changed
• No tool-generated co-author trailers or footers