Expose ordered engine support CSS as its own asset

[borkweb]

Why

HtmlTransformer::materializeAuthorStylesheet() concatenated two different things into one
generated asset (source: 'author-css'): the engine support CSS the transformer generates so
its own output works, and the rewritten author stylesheet. A consumer that writes its own
author CSS cannot take the second, and had no reliable way to extract the first.

The real consumer — automattic/site-build — was reduced to grepping the asset for
/\.(be-inline-geometry-[a-z0-9-]+)\{[^{}]*\}/ and discarding everything else. Measured across six
generated sites, every engine marker class appeared in the block markup with zero matching CSS
in the theme:

• <mark> richtext carriers rendered with the browser default yellow highlight on headlines and
  brand text — 15–42 unstyled marks per site, an accessibility regression.
• Synthetic <p> wrappers kept UA margin:1em 0, so header/nav lockups sat misaligned.
• blocks-engine-control-* button rules were dropped — 10 occurrences in one site's markup, 22 in
  another, 0 matching rules in either theme. Transformed buttons rendered unstyled.

A shape-based regex in the consumer was considered and rejected: the generated rules already use four
distinct selector shapes (:where(.blocks-engine-*), :where(mark)[style*=…],
.blocks-engine-control-X.blocks-engine-control-X>.wp-block-button__link, .wp-block-navigation…),
so any regex under-carries. The boundary is knowledge only the transformer has.

What changed

1. The generated stylesheet is split by ownership. author-css now holds only the rewritten
author stylesheet plus its own leading at-rule preamble. Engine-generated rules move to
source: 'engine-support' assets.

2. Cascade order is preserved and expressible. Some support rules deliberately preceded the
author CSS (carrier rules keeping inline precedence) and others deliberately followed it
(list-navigation display:flex!important, mobile overlay, button rules). A single support asset
would have lost that, so assets carry stylesheet_placement: before-author | author |
after-author. A consumer that never takes the author CSS at all still knows which side each rule
belongs on. ArtifactCompiler orders them beforeAuthor → manifest → other → afterAuthor.

3. New blocks-engine-css-owned-layout neutralizer. Rules already existed for
css-owned-flow, css-owned-grid and css-owned-layout-item, but nothing for css-owned-layout.
In a real WordPress theme a layout:constrained block gets core CSS applying
max-width: <contentSize>; margin-inline: auto to its non-aligned children — which collapses a
CSS-owned flex container toward the centre. Observed downstream: a primary nav rendering ~860px wide
and centred inside a header the author CSS sized at 1280px. The static-site output never sees core
layout CSS, so the gap was invisible there.

4. The richtext-marker <mark> is now self-sufficient. It carries inline
background-color:transparent;color:inherit, matching what the styled-inline path already did — the
divergence between the two paths was itself the bug. The :where() stylesheet reset is retained,
not replaced: zero specificity means it never fights the inline value, and it still covers consumers
whose sanitizers strip inline styles.

Verification

Gate	Result
composer test	exit 0
Parity fixtures	275 passed
tests/unit/engine-support-css-asset.php	49 assertions

The 11 modified parity fixtures are legitimate expected-output changes: marks gain inline
background-color:transparent, and one asset count goes 1 → 2 where the new css-owned-layout
neutralizer applies.

VERSION → 0.4.17; CHANGELOG names the new asset source, the placement enum, and the neutralizer.