[ellenbauer]

Refs #2726. Follow-up to #2744.

What was wrong

Review of #2744 found that a projected captured modal can still open invisible, or keep the browser's dialog box.

1. Id rules never matched the dialog during analysis. The dialog has a generated id, and fix(php-transformer): keep the source panel box on a projected modal dialog #2744 carries the source panel id as its editor anchor class. Id selectors reach the dialog in the output CSS through that class, but the source analysis (selector matcher and candidate indexes) still only compared the id attribute. So CapturedDialogConverter::modalPlacement(), which reads the cascade, missed a panel positioned only by id rules (#panel{position:fixed;top:0;left:10px;width:200px}), and the dialog kept the browser box.
2. A rest rule that hides the panel dropped the dialog block. HtmlCompilation::sourceElementStartsHidden() read display:none at rest as hidden content, and the dialog was not emitted. This was already true for class rules on trunk (.panel{display:none}), and it would have become true for id rules once they match.
3. Without a captured ancestor state, the rest rule won on the open dialog. With #panel{display:none;transform:translateX(-100%)} and body.nav-open #panel{display:block;transform:none}, the open dialog stayed hidden and off canvas. The existing closed-state repair (ClosedStateNormalizer) does not cover this. It only reacts to :hover and :focus reveal states, its repair stylesheet is editor-only, and it exempts positioned layers.

What changed

• CssSelectorMatcher::matchesCompound(), CssSelectorMatchCache::styleRuleCandidates() and AuthorStyleAnalysis::selectorCandidates() and selectorCanMatch(): an element carrying an id's editor anchor class is treated as that id's holder. Only a projected captured dialog carries such a class in source HTML.
• HtmlCompilation::sourceElementStartsHidden(): a captured <dialog>'s own rest state is not hidden content; the dialog block owns its closed state.
• StyleResolver::capturedDialogOpenStateValues(): walks the raw author rules (the static analysis drops transform and drops rules that match no element yet). A state rule is an unconditioned rule whose rightmost compound matches the dialog but whose full selector does not. For display, visibility, opacity and transform, where the rest cascade hides or offsets the panel and a state rule reverses it, the state rule's value is returned. A transform the open state keeps, such as a centring translate, is not touched.
• CapturedDialogConverter: a modal carries those values in its inline style, without !important. That beats the rest rules, and the block stylesheet's dialog[data-blocks-engine-triggers]:not([open]){display:none!important} still closes it.

Tests

• New tests/unit/captured-dialog-open-state.php, 6 assertions, compiled through ArtifactCompiler with a wired copy holding #panel:

  • id rest state: block kept, style="display:block;transform:none", placement source;
  • class rest state: style="display:flex;opacity:1";
  • id-only positioning: placement source and no inline values;
  • centring transform kept by the open state: left alone.

  On trunk 4 of 6 fail.

• captured-dialog-source-panel-box, captured-dialog-projector, captured-dialog-closed-visibility, captured-dialog-trigger-survives, captured-dialog-dropdown-panel and css-selector-matcher still pass. The layer direction contract passes, and phpstan is clean on the touched files.

• capture-evidence-equivalence: 349/349 unchanged.

• Full unit run (memory_limit -1, one process per test): 429 pass, 0 fail.

Review round 2 (second commit, 59a9430)

• State rule selection. The last state rule in source order used to win: .dark-theme #panel{opacity:.5} after body.nav-open #panel{opacity:1} gave opacity:.5, and .no-js #panel{display:block} replaced display:flex.
  • When the capture recorded an opened ancestor class (data-blocks-engine-ancestor-state), only state rules naming it are trusted. If none names it, nothing is carried.
  • With no recorded state, every state rule must agree on a property, or nothing is carried for it.
• Transform offsets. translate(-50%,-50%) counted as a rest offset, so .no-js #panel{transform:none} un-centred a modal that is fine on trunk. Only a translation of 100% (a full panel) or 100vw or 100vh or more along an axis now counts as moving the panel out of view.
• Orphaned dialog. A dialog none of whose triggers can ever show (the trigger or an ancestor is display:none at its base with no conditional value that shows it, or carries hidden) is dropped again. A burger hidden on desktop but shown in a phone media query keeps its dialog.

Dialog probe matrix against trunk 2a963f0 (the reviewer's dlg2.sh cases plus two trigger cases):

Case	Trunk	This branch
id rest none + body.nav-open	no inline values	display:block;transform:none, placement source
opacity: nav-open 1, then dark .5	none	none (disagree), placement source
display: nav-open flex, then no-js block	none	none (disagree)
centred translate + no-js none	none	none
rest only under @media	none	none
class rest none + .menu-open	dialog dropped	display:block
#menu:hover ~ #panel	none	none
no rest hide	no placement	placement source
trigger display:none everywhere	dialog kept (rest id rule unseen)	dialog dropped
trigger hidden on desktop, shown on phone	kept	kept, display:block

captured-dialog-open-state.php now has 12 assertions: 7 fail on trunk, and 5 fail on the previous PR head (the review cases). capture-evidence-equivalence: 349/349 unchanged. Full unit run (memory_limit -1, one process per test): 429 pass, 0 fail.

Import proof (workbench, new-vocabulary captures)

Studio CLI create --from with the liberation workbench captures (DLA main 569c336). The baseline is SSI v1.31.2 with BE trunk 2a963f0 (wb-*-trunk). The proof package is the same ZIP with #2780, #2781 and #2783 merged on 2a963f0 (ssi-be-proof-2726-combined-e51205f.zip). iPhone user agent at 375; desktop at 1280. Quality gate passed on every import (quality_pass: true, fail_import: false). The proof sites are stopped, not deleted.

• Dunbar Oaks (localhost:8953): the menu dialog keeps placement source with no inline open values (its panel has no state rule to replay). Box at x=27.5, 320px wide, beige layer, the same as trunk (localhost:8952).
• Dunbar at 1280 and at 375 with the menu closed: 0 differing pixels.
• St Francis (localhost:8955): the dialog still opens through its bound trigger with all root items on screen, the same as trunk.
• None of the workbench sites uses an id-only rest state, so the new cases are covered by the compiler probes and unit tests above, not by these imports.

🤖 Generated with Claude Code

[ellenbauer]

Please hold: our independent review found the open-state picker takes whichever state rule comes last (e.g. a .dark-theme or .no-js variant), and a centring translate(-50%,-50%) is treated as an at-rest offset, un-centring modals that were fine on trunk. Fix in progress on this branch.

[ellenbauer]

Hold lifted. Merged after three independent review rounds. Follow-up to #2744: id rules match a projected dialog through its editor-anchor class, so source placement sees id-only positioning; a dialog's own rest state no longer drops it; open-state values (display, visibility, opacity, out-of-view transforms) are carried inline only from state rules naming recorded ancestor state, or that all agree; centring transforms are left alone; a dialog whose triggers can never show is dropped. Known gaps, same as trunk: pixel-offset drawers, JS-added trigger classes. Refs #2726. cc @chubes4