fix: builder layout meta stays raw, never modelled as content (too_large on deep Elementor trees)

[ABB65]

Problem

On the Elementor golden, the migration export failed as a whole: EXPORT_FAILED too_large with "Content nesting exceeds the supported modelling depth". Deep _elementor_data trees on the landing and about pages were modelled into meta_* content fields and wp-structured-values rows, and one of them went past the modelling depth.

Change

• Page-builder meta (Policy::BUILDER: _elementor_*, _et_pb_*) is no longer modelled as content. It is kept only in the raw evidence (raw-posts / RawIR), where Migrate reads it. This matches @contentrain/wp-import, which never turns underscore meta into content fields, so status and entry parity do not change.
• Past the modelling depth, value() stops throwing. It keeps the rest of the subtree as JSON text and records a nesting-too-deep: kept as JSON text warning. A single deep value can no longer fail the export.

Verification

• tests/builder.php: 64 lines of new checks.
• The full acceptance suite passes on an isolated Docker stack, including parity with wp-import 0.5.7.
• Elementor golden, with this build installed:
  • The export completes: 8 models, 13 entries, 53 files.
  • It has no nesting warnings and no issues.
  • _elementor_data is kept in rawir.json.
  • wp-page has no meta_elementor_* fields.
• ACF golden: the export completes, with 15 models and 20 entries.

Open points (separate work)

• The comment on Policy::BUILDER still mentions CORE_META. wp-import actually skips every _ key.
• Bridge still models other underscore meta (_yoast_*, _aioseo_*). This is a separate parity gap with wp-import.
• The release and the WordPress.org plugin update are out of scope for this PR and need the founder's approval.

[ABB65]

ONAY bcc8766 (t9)

No silent loss:

• Builder meta that is skipped (only the 10 exact keys in Policy::BUILDER) isn't lost. raw-posts gets the whole $p (models.php:271), so the keys stay in RawIR.
• The test checks it end to end: the 12-container tree's leaf Golden deep can be read from raw-posts, and _et_pb_use_builder stays in raw-posts too.
• Past the depth limit, the subtree is kept as JSON text and exactly 1 nesting-too-deep warning is written. The gap is visible, not silent.

Backward compatibility: migrate (main) reads _elementor_data only from raw/RawIR (facts behaviors.ts/builder.ts). No meta_elementor_*/meta_et_pb_* content field is consumed in migrate or studio (git grep: 0). The REST envelope doesn't change; only these fields drop out of the content model.

PHP/WP: the plugin needs PHP 7.4 / WP 6.2. php:7.4-cli php -l on the plugin + includes + tests/builder.php: 0 errors. The new code uses no API newer than 7.4 (preg_match, wp_json_encode, ??).

Local gates (Docker, isolated stack, torn down afterwards):

• tests/run.sh acceptance: exit 0, 378 PASS. The FAIL matches in the log are only test names ("fails validation") and wp-cli cache warnings.
• tests/builder.sh: 26/26 PASS, exit 0.
• tests/check-plugin.sh (Plugin Check 2.1.0): 0 errors, 14 warnings, all from files that already use $wpdb (DirectQuery/NoCaching, the ones documented in RELEASING.md). The file changed in this PR produces none.
• npm test 2/2 · test:secrets 85 files, 0 secrets · test:reader ok.
• CI: PHP 7.4/8.3 syntax, metadata, small/large, acceptance all ✓. merge-tree with main is clean.

Non-blocking notes:

1. tests/builder.php prints "Builder checks: 0" (the global $checks counter doesn't survive the eval-file scope). Old behaviour, cosmetic; the PASS lines are correct.
2. If a user lists one of the BUILDER keys in selected_meta, it now produces no content field and no warning (it's still in raw). An edge case; a "builder meta stays raw" warning could be added.
3. wp_json_encode returning false would give ''. The warning is still written, so it's not silent.