test(e2e): verify in a real browser that a dialog hides the live regions (#501)

[mforce]

Part of #501. That issue stays open — this lands the half a machine can check; the screen-reader pass is still owed.

Why

#485's premise — that #483's inert sweep takes the offscreen announcers out of the accessibility tree, so an announcement made then is never heard — was only ever inferred from the source. jsdom implements neither live regions nor the inert IDL, so nothing in web/ could observe it.

What it proves, in real Chromium

1. Both announcers are mounted, empty, exposed and in the accessibility tree on a healthy screen, with no live role.
2. Opening a dialog removes them from the tree while leaving the DOM and their attributes untouched. Control in the same snapshot: the dialog's own controls stay.
3. Closing it returns them.
4. A standing farm warning is carried by the visible role="alert" banner, and the announcer is never written to across two dialog cycles.

CDP is not a style choice

Playwright's own accessibility surface does not model inert — measured, not assumed: ariaSnapshot() still lists an inert node, isVisible()/isEnabled() still return true, only click() notices. The same spec written with getByRole would pass whether or not the app inerts anything. src/ax.ts is the sanctioned reader.

Three traps, each of which produced a false result first

• The sweep runs in a passive effect, so waiting for the dialog to appear does not settle it. Every transition now waits on the sweep, which is also an assertion.
• A target and its control must share one tree snapshot — a control read several round trips later cannot rule out a tree that moved. That is what ax.nodes() is for.
• "Stayed empty" is a non-event, so a snapshot read only says empty when I looked. A MutationObserver records every write for the whole test.

Also recorded, so the manual pass has less to establish

Both of #501's deferred designs rest on a browser assumption nobody had checked. aria-live="off" does suppress the implicit politeness of role="alert" (paired with a control, because a non-live element reports the same null), and un-inerting one subtree while a dialog is open does return it to the tree. The Chromium build is annotated — the CDP accessibility domain is experimental.

Mutation

The harness's first DOM-level mutants; src/mutants.ts records what that instrument is and is not faithful to. The three announcer mutants differ only in when they write, and each is the only one that reaches its assertion — the first version had one mutant covering a third of the test it named, and the harness reported a clean kill and said nothing.

Full run: baseline GREEN, 15 killed, 0 survivors, restore GREEN.

Not covered, stated in the spec header

• No screen reader is involved. Tree presence is the precondition, not the utterance.
• The missed-then-delivered path — the core of fix(web): announce the update and farm warnings a dialog made inert #499's fix — has no product affordance to drive it: the update banner needs a second service worker to park in waiting (Playwright cannot provoke it), and the farm warning needs /account to fail on a re-read, whose only callers are the banner's own retry and the settings save, neither of which runs with a dialog open. A test-only trigger was considered and rejected.

Both are docs/runbooks/screen-reader-verification.md's job.

Review

Four local passes before pushing (pi + three agents: adversarial, conventions, flake/CI). They found the passive-effect race, the uncovered anti-nag loop, a false claim of mine in mutation-check.sh, an undelivered "same snapshot" invariant, and the snapshot-vs-cumulative weakness. One reported finding — aria-hidden slipping past inTree — did not reproduce: Chromium 151 drops those nodes outright, same as inert.