Prune accessibility-ignored empty nodes

[morisil]

Summary

toSemanticEvents(respectAccessibility = true) skips subtrees excluded from the accessibility tree by aria-hidden="true", display:none, and visibility:hidden (isAccessibilityHidden). Chrome's tree drops more than that: empty inline wrappers, hidden-attribute elements, elements with no accessible content, some content clipped/positioned offscreen specifically for assistive tech. A node Chrome ignores and that carries no text is safe to drop, removing structural noise the current capture keeps.

Depends on #48 (shared introspection layer). Lowest priority of the accessibility set — hold until a real capture demonstrates the noise; do not implement speculatively.

Signal

An element is ignored by the accessibility tree roughly when its computed role resolves to nothing renderable AND it contributes no name/text. Detection options:

• computedRole == "" (present-but-roleless / ignored) combined with an empty computedName and no descendant text. The combination is essential — see the trap.
• The hidden attribute and inert are cheap explicit signals not currently checked by isAccessibilityHidden; those can be added independently of the flag.

⚠️ The trap

This is the most dangerous of the accessibility issues because "ignored" is easy to over-apply:

• computedRole == "" also describes a layout-table node (the shipped layout-table case) and presentation nodes (Generalize role=presentation/none unwrapping beyond tables #50) whose children are meaningful — those must be unwrapped, not skipped. Skipping would delete content.
• Offscreen "visually-hidden but screen-reader-visible" patterns (.sr-only) are the OPPOSITE — present in the tree, must be kept.

So the rule must be ignored AND empty (no text, no accessible name, no meaningful descendants) → skip; never "ignored" alone. The classifier ordering from #48 matters: unwrap/rename cases must be evaluated before the skip case.

Suggested approach

Add a Skip case to the #48 classifier, gated on the conjunction above, evaluated after the unwrap/rename cases. Start conservative: only hidden/inert attributes + truly empty roleless nodes. Expand the net only against evidence from rendered-Markdown diffs on real dumps (./gradlew :markanywhere-html:renderDumpMarkdown).

Test approach

Deterministic, under ChromeHeadlessExperimentalA11y:

• <span></span> between content → dropped, surrounding text intact.
• <div hidden>secret</div> → dropped.
• Negative / safety: <div class="sr-only">screen reader text</div> (offscreen but in the tree) → KEPT.
• Negative / safety: a roleless wrapper with text children → children survive (not skipped).

Out of scope

• Anything that moves or reorders content (aria-owns).
• Re-deciding the existing aria-hidden/display:none/visibility:hidden behaviour, which already works.

[morisil]

Rescope — unblocked from #48 (closed), still a (low-priority) live feature.

Architecture changed (see #48): no browser-side flowElement Skip classifier / karma launcher. Today only ignored images are pruned — CapturePage.kt records AccessibilityAnnotations.IGNORED = "true" for an <img> whose axNode().ignored is set, and ApplyAccessibility.kt drops it. The general "ignored AND empty → skip" case is not implemented, which is fine: this issue itself says hold until a real capture demonstrates the noise.

New shape, if/when pursued:

• The capture side already has the AX verdict in hand (axNode()?.ignored); extend the IGNORED annotation beyond images (currently restricted to img because most other ignored nodes are structurally meaningful), OR add a cheap hidden/inert attribute signal at consume time.
• Add a Skip case to the ApplyAccessibility operator, ordered after the unwrap/rename cases (Recover semantic structure from ARIA roles (div-soup → headings, lists) #49/Generalize role=presentation/none unwrapping beyond tables #50), gated on the conjunction ignored AND no text AND no accessible name AND no meaningful descendants — never "ignored" alone.
• The ⚠️ trap is unchanged and is the whole reason this is lowest priority: a roleless node with meaningful children must be unwrapped (Generalize role=presentation/none unwrapping beyond tables #50), not skipped; .sr-only offscreen text is in the tree and must be kept.

Tests move to ApplyAccessibilityTest (deterministic). Keep this parked until ./gradlew :markanywhere-html:renderDumpMarkdown on a real dump shows the noise it's meant to remove.