Refine runtime textarea workspace detection

[chubes4]

Summary

• distinguish runtime textarea editors from real forms and structural pseudo-forms
• preserve only the tight JavaScript-owned editor island while keeping surrounding editorial layout native and editable
• cover real forms, pseudo-forms, standalone editors, nested editors, and unrelated runtime targets with parity fixtures

This addresses a generic transformer failure found while investigating Automattic/static-site-importer#623 under Automattic/static-site-importer#489: a runtime-targeted textarea could cause an ordinary contact or editorial wrapper to be preserved wholesale as core/html.

Implementation

hasWorkspaceSurface() now treats a textarea as workspace evidence only when it is runtime-bound and is not part of a real form. Structural pseudo-forms remain on form semantics unless the tight pseudo-form candidate contains an internal runtime-bound non-control target, such as a preview. That editor evidence applies only to the tight candidate, so outer editorial ancestors remain native blocks.

How to test

1. Check out this branch on a fresh clone.
2. Run composer --working-dir=php-transformer test.
3. Confirm the full transformer suite passes, including 235 parity fixtures.
4. Run php -l php-transformer/src/HtmlToBlocks/HtmlTransformer.php.
5. Confirm the syntax check reports no errors.

Validation

• composer --working-dir=php-transformer test passed.
• PHP transformer parity fixtures passed: 235 fixtures.
• php -l php-transformer/src/HtmlToBlocks/HtmlTransformer.php passed.
• git diff --check passed.
• An SSI control/candidate visual matrix over six representative fixtures completed with unchanged mismatch rates, native conversion rates, fallback ratios, editor validity, and failure categories. The only delta was one additional acceptable form-graft diagnostic caused by the tighter boundary.
• A full ArtifactCompiler comparison compiled all 383 HTML entrypoints in the Blocks Engine fixture corpus with each site's HTML, CSS, and JavaScript bundle. 363 outputs were byte-identical. The 20 changed outputs were all textarea/form boundary cases.
• Across those 20 outputs, the candidate exposed 683 additional native blocks and reduced opaque core/html content by 77,558 bytes, with no aggregate increase in core/html block count.
• In 19 outputs, a broad section-level app shell became a bounded form-level island. The 20 newly visible fallback records describe content that was already hidden inside those broad app shells: 19 runtime forms and one unsupported progress element.

Compatibility

This intentionally changes runtime-island classification for runtime-bound textareas. Real forms and structural pseudo-forms retain bounded form/control preservation, while genuine textarea editors retain bounded app-shell preservation. No public APIs or extension paths change.

AI assistance

• AI assistance: Yes
• Tool(s): OpenAI GPT-5.6 Sol with OpenCode
• Used for: Investigating transformer evidence, implementing the classification boundary and parity fixtures, and reviewing regression cases. Chris Huber reviewed and remains responsible for the change.