Make visual-compare diagnostics region-aware and semantic

## Problem

`wordpress.visual-compare` captures the first 160 visible DOM elements, matches source and candidate by exact generated path, omits pseudo-elements, and ranks rectangle overlap rather than mismatched pixels. Gutenberg wrapper changes therefore turn corresponding elements into added/removed rows and late-page mismatch regions lose actionable element attribution.

Observed in Static Site Importer fixture `15-saas`: only 2 flow elements matched while 158 were reported added and 158 removed; both 160-element snapshots were truncated. The pixel diff clearly showed navigation, button, text, and pseudo-element paint drift that the explanation did not rank explicitly.

## Required outcome

- Match elements semantically when wrappers and `:nth-of-type()` paths differ, with deterministic confidence and ambiguity handling.
- Capture visible `::before` and `::after` paint/style evidence under their host element.
- Augment bounded snapshots with elements intersecting top mismatch regions, including late-page elements.
- Rank attributed elements by mismatched-pixel impact rather than rectangle overlap alone.
- Preserve explicit selector capture and bounded output with exact selection/truncation metadata.

## Acceptance

- Wrapper insertion does not collapse semantic matching.
- A pseudo-element paint mismatch is named in the explanation.
- An element after the initial document-order budget is captured when it intersects a top mismatch region.
- A large unchanged wrapper ranks below a smaller element covering actual mismatch pixels.
- Deterministic tests cover all behaviors.

Start at the wordpress.visual-compare entry point and inspect its snapshot, matching, mismatch-region, and explanation logic, then review the existing tests. Define deterministic coverage for semantic wrapper matching, pseudo-elements, late-page region capture, pixel-impact ranking, and selector/truncation metadata. Done means all listed acceptance behaviors pass without losing bounded output.

Written by the indexing model from the issue text.