Convert StyleResolutionTrait into a StyleResolver collaborator (#242)

[chubes4]

Slice 6 for #242, workstream 1. Follows #1310, #1312, #1315, #1316, #1317.

This is the big one. StyleResolutionTrait was 2,822 lines — the largest single-consumer trait mixed into HtmlTransformer, and the single largest contributor to that class's object scope.

The surprise

I expected this to be the hardest remaining trait. Measured, it was the easiest:

trait	lines	external deps
StyleResolutionTrait	2,822	15
FormDispatchTrait	1,865	30

The biggest trait had half the external coupling of one two-thirds its size. Size and coupling were inversely related here, which is worth knowing before picking the next one.

What changed

StyleResolutionTrait → StyleResolver, behind a 15-operation StyleResolutionContext. Per-transform state (author styles, source styles, layout geometry, presentation cache, transformation evidence) resolves through closures, because the resolver is built once with the transformer but must see the state of the transform currently running.

StyleAttributeMapper and HighValueStyleBoundaryPolicy stay self-owned lazy singletons — they were already the resolver's own, not the transformer's.

292 call sites migrated, trait deleted, no shims:

	call sites
HtmlTransformer	226
SvgMaterializationTrait	18
ElementConversionTrait	18
NavigationStyleProjectionTrait	13
FormDispatchTrait	13
NavigationToggleSuppressionTrait	2
DomHelpersTrait	2

A trap this surfaced

DomHelpersTrait is shared by four classes, only one of which is the transformer. Its canonicalizeLinkUrls() invalidated a selector-match cache that only the transformer owns — and it is reached transitively from innerHtml(), innerHtmlPreservingWhitespace(), and outerHtml(), which all four consumers call.

A naive migration would have left the shared trait referencing a collaborator that SubtreeClassifier, FallbackEmitter, and SemanticParityReporter do not have. It would have worked in the corpus run and blown up on whichever path first canonicalized a link from one of those three.

Fixed with an onSourceMarkupMutated() hook on the trait that defaults to a no-op; HtmlTransformer overrides it to invalidate the cache. The shared trait no longer knows about a collaborator three of its consumers lack.

Behavior preservation

385 fixtures — DIFFERING: 0 — threw: 0

The corpus check earned its keep twice here: it caught a namespace collision on HtmlTransformerAnalysisCache (resolved into ...\HtmlToBlocks\Style\ instead of ...\HtmlToBlocks\) as 385/385 TypeErrors the moment it was introduced.

composer test exit 0.

Test updates

Three unit tests reflected on style methods via HtmlTransformer. They now reach StyleResolver through the transformer's collaborator. presentationCacheKey stayed private and is reached by reflection on the resolver rather than widening its API for a test.

Impact

	trunk	after
HtmlTransformer.php	12,684	12,721
trait lines on the transformer	8,956	6,146
effective object scope	21,640	18,867

−2,773 lines off the object scope in one slice — more than the previous five slices combined (−386).

Cumulative since the audit: effective object scope 22,026 → 18,867 (−14%), three traits eliminated, convertElement() 441 → 301.

Scope

Five single-consumer traits remain, 5,542 lines: FormDispatchTrait (1,865), NavigationStyleProjectionTrait (1,319), SvgMaterializationTrait (983), NavigationToggleSuppressionTrait (781), ElementConversionTrait (594).

Given what this slice showed about size vs coupling, the next one should be chosen by measuring external dependencies rather than line count.

---

AI assistance disclosure: implemented and drafted by Claude Sonnet 4.6 running in Claude Code, operated by @chubes4. The AI measured trait coupling to select the target, captured the pre-change corpus baseline, converted the trait, migrated all 292 call sites, identified and fixed the shared-trait hazard in DomHelpersTrait, and verified byte-identical serializedBlocks and fallback counts across 385 fixtures. Reviewed by a human before opening.