Keep multi-child hero stacks grouped

[chubes4]

Summary

• Tightens ColumnsPattern so split-layout heuristics only emit core/columns for actual two-pane structures.
• Keeps multi-child content stacks such as .hero-content as core/group with the source class preserved, allowing the authored stylesheet to control flow.
• Adds regression coverage for the seafood-style hero content stack.

Evidence

• Issue: Closes C1 visual parity: style materialization gap — 97-100% native fixtures render 76-89% wrong (12-portfolio, 21-seafood-popup) #486
• Baseline c1 ratios from SSI matrix (8421204 / transformer ebc3997): 12-portfolio 0.889, 21-studio-web-seafood-popup 0.759.
• Single-fixture proof with this branch and current wp-codebox main:
  • 12-portfolio: 0.087 visual mismatch, native 0.966, fallback 0.035.
  • 21-studio-web-seafood-popup: 0.120 visual mismatch, native 1.000, fallback 0.000.
• Full c1 proof completed with 7/7 fixtures passing hard gate. The zero-threshold pixel ratios remain unstable for subtle full-page background/text differences (e.g. visually aligned seafood screenshots report high ratios in multi-fixture runs), so the PR is intentionally scoped to the concrete generic layout misclassification fixed here.

Tests

• php php-transformer/tests/unit/block-style-support-conversion.php
• php php-transformer/tests/unit/corpus-detectors.php
• php php-transformer/tests/contract/run.php
• Direct SSI single-fixture matrix runs for 12-portfolio and 21-studio-web-seafood-popup passed.
• Direct SSI c1 matrix run with --max-complexity 1 passed 7/7 hard gate.

AI assistance

• AI assistance: Yes
• Tool(s): GPT-5.5 via opencode subagent, orchestrated by Claude (claude-fable-5)
• Used for: Diagnosis, implementation, tests, and PR drafting; reviewed by Chris Huber