Preserve auto-fit grids, inline flex containers, and authored button font-family

[borkweb]

Three defects that all change rendered geometry. Each was confirmed against real renders before any code changed, and each has a committed RED test that failed at the test-only commit.

Problem 1: auto-fit is converted to a grid layout WordPress renders as auto-fill

StyleResolutionTrait::autoRepeatMinimumColumnWidth() accepted repeat(auto-fit|auto-fill, minmax(W, 1fr)) and converted BOTH to a native "layout":{"type":"grid","minimumColumnWidth":"W"}.

wp-includes/block-supports/layout.php hardcodes auto-fill in every branch that renders minimumColumnWidth, so the attribute cannot express auto-fit at all. The two differ in rendered geometry: auto-fit collapses tracks left empty, auto-fill retains them. An authored auto-fit grid therefore rendered with the empty tracks kept and the real content squeezed into part of the measure.

Measured on a design fixture with three affected sections, design tracks versus delivered tracks:

Section	Design (authored)	Delivered before	Delivered after
minmax(260px), 2 items	repeat(auto-fit,minmax(260px,1fr)) = 592 592 0 0	minimumColumnWidth:260px, rendered as auto-fill = 267 267 267 267	authored track list, verbatim
minmax(270px), 3 items	repeat(auto-fit,minmax(270px,1fr)) = 389 389 389 0	minimumColumnWidth:270px = 284 284 284 284	authored track list, verbatim
minmax(280px), 2 items	repeat(auto-fit,minmax(280px,1fr)) = 586 586 0	minimumColumnWidth:280px = 366 366 366	authored track list, verbatim

Across that whole page: native minimumColumnWidth attributes 5 to 0, is-layout-grid classes 10 to 0, delivered auto-fit track lists 0 to 5. A fourth grid on the page has 4 items, where auto-fit and auto-fill agree, which is why only three sections were visibly wrong.

Control group: a second design fixture authors minmax(min(100%,14rem),1fr), which never matched the conversion regex. Its 6 auto-fit track lists are byte-identical before and after this change.

Problem 2: inline display:flex is dropped while the block stays marked css-owned-flow

cssOwnedGroupAttributes() drops the native layout attribute when an author flex or grid container is demoted to a css-owned core/group. Grids survive because CSS_OWNED_GRID_CARRIER_PROPERTIES rides their inline declarations to the generated stylesheet. There was no flex equivalent, so a block marked blocks-engine-css-owned-flow lost the declaration that made it a flex container, and its children stacked.

A footer in one design fixture, before and after, same carrier class:

before: .be-inline-geometry-<hash>{max-width:var(--wide-size) !important}
after:  .be-inline-geometry-<hash>{max-width:var(--wide-size) !important;display:flex !important;
        flex-wrap:wrap !important;gap:clamp(1.8rem,5vw,4rem) !important;justify-content:space-between !important}

No display rule existed anywhere in the generated stylesheet before. Three such containers on that page alone.

Problem 3: core/button drops authored font-family

ButtonStyleResolver::BUTTON_TYPOGRAPHY projected fontSize, fontWeight, letterSpacing, lineHeight and textTransform onto buttons and dropped font-family. core/button registers a fontFamily attribute (packages/blocks-engine/src/core-block-attrs.json, generated from @wordpress/block-library 9.42.0), which core injects only when the typography fontFamily support is enabled, so this was a dropped declaration rather than an unsupported one.

The authored .btn class is consumed into block attributes, so its surviving rule is rewritten behind :where() at zero specificity and matches nothing that can win. theme.json's styles.elements.button.typography.fontFamily then substituted the theme typeface, widening hero CTAs enough to wrap them onto two lines.

Buttons on one design fixture, before and after:

before: 8 buttons, no fontFamily attribute at all
after:  8 buttons carrying "Trebuchet MS", "Segoe UI", system-ui, -apple-system, Roboto, "Helvetica Neue", Arial, sans-serif

This has to be fixed per authored declaration rather than by a blanket downstream change: across seven design fixtures the .btn family matches the body family in three, and differs in four. One fixture whose buttons are correct today only because the theme preset happens to match now carries its authored stack explicitly ("Archivo", ..., its own heading stack, resolved), so it stays correct rather than being correct by coincidence.

Fix

1. autoRepeatMinimumColumnWidth() now matches auto-fill only. auto-fit returns '' and falls through to the same css-owned-grid carrier that already holds every other track list WordPress cannot express, so the author rule stays the single owner of the exact authored track list. auto-fill keeps converting natively.
2. CSS_OWNED_FLEX_CARRIER_PROPERTIES, isCssOwnedFlexElement() and cssOwnedFlexAttributes() mirror the grid carrier for flex containers. An inline display that overrides class-owned layout is skipped, because Preserve inline display overrides through the geometry carrier #883 already carries those properties complete at the non-important specificity tier; forcing them would move them to the important tier and change that output.
3. fontFamily joins the button typography subset. A raw authored family is a custom value rather than a preset slug, so it rides in style.typography.fontFamily. BlockFactory's typography map did not map that attribute either, so the block attribute alone never reached the rendered link; the serializer now emits font-family inline the way core's style engine does. This part is deliberately broad rather than button-scoped: the map is the generic serializer, and the drop affected any block that mapped the declaration.

Inline style outranks theme.json element styles, so no !important bridge was added for the font-family path.

Verification

Local, on this branch rebased onto c0d66eafe:

• composer test:unit exit 0, 39 scripts, 0 failures.
• composer test exit 0, 51 scripts, 0 failures, 275 parity fixtures.
• New contracts: auto-fit-grid-carrier.php 11 assertions, css-owned-flex-carrier.php 12, button-font-family-carry.php 9.
• RED evidence at the test-only commits: 6 failed / 5 passed, 6 failed / 6 passed, and 5 failed / 4 passed respectively.
• block-style-support-conversion.php 130 assertions after resolving the three-way append collision as a union. Trunk's own copy of that file reports 129 assertions, so the union is trunk's 129 plus the one assertion added here, with neither side dropped.

Seven parity fixtures changed, each because it encoded one of the defects rather than to make a test pass. Three asserted the auto-fit to native conversion, one of which was renamed because its name asserted the buggy contract. Four asserted flex output in which the authored display:flex had already been dropped, including one whose own description claimed a horizontal layout it was not getting. No parity fixture changed for the font-family fix.

Rendered-class inventory across the 89-site corpus, comparing this branch against c0d66eafe: zero has-*, is-layout-*, wp-block-* or wp-element-* class tokens are ADDED anywhere. Every class-token change is a removal of is-layout-grid / wp-block-group-is-layout-grid on containers that should never have had core grid layout. This is the specific failure shape from #885, where has-border-color triggered core's all-sides border-style:solid and painted phantom borders; nothing here emits a class that makes WordPress paint something the author never wrote.

Solved-gate fixtures, transformed output compared against c0d66eafe:

• 89-static-site-importer-architecture: byte-identical.
• 15-saas: differs only by 8 rendered font-family declarations on buttons, value 'IBM Plex Mono', 'Courier New', monospace. That fixture's source authors .btn { font-family: var(--font-mono) } with --font-mono: 'IBM Plex Mono', 'Courier New', monospace, so the delivered typeface now equals the source's. No auto-fit, flex-carrier or is-layout-grid change on either fixture.

The Solved Site Promotion Gate cannot be run locally, since it lives in another repository and needs a WordPress rendering rig. It is the gate that matters most here, because all three defects change rendered geometry and static parity compares output structure rather than pixels. It is left to CI on this PR to arbitrate at the 0.00% threshold.

Static style parity, this branch against c0d66eafe

Run directly rather than through composer, against a baseline worktree whose vendor/ was copied rather than symlinked, since a symlinked vendor makes Composer's $baseDir resolve to the other tree and silently invalidates the comparison. Both trees were checked by reflection before the runs to confirm they resolved to their own sources.

Not byte-identical. Across 87 fixtures: mean score 0.664049 to 0.667159 (+0.003109), total findings 12385 to 12313 (-72), no status transitions. 23 fixtures moved: 16 improved, largest 69-markdown-editor 0.2704 to 0.5039 with findings 173 to 116, and 7 regressed between -0.0028 and -0.0012.

All 7 regressions are font-family findings, and all are comparator artifacts rather than fidelity losses:

• In 4 of them the source side reads var(--X) while the candidate reads the literal value of --X. Each variable was resolved by hand and is byte-equal, for example --sans: 'Space Grotesk', system-ui, sans-serif against a candidate value of 'Space Grotesk', system-ui, sans-serif. The comparator does a string comparison and cannot resolve var().
• In the other 3 the findings land on a <span> inside a button. Each of those fixtures' .btn rule does author the family the candidate now carries. The comparator resolves the inner span's inherited family to the document default instead of inheriting from .btn, so the candidate is the more faithful side.

The pixel-level check that this signal cannot provide is the Solved Site Promotion Gate, which is why that gate is the one that decides here.