Separate required HTML validation outcomes from detailed reports

[chubes4]

Summary

Continue #1360 by separating required HTML validation facts from detailed report maps at the diagnostics-consumer boundary.

• HtmlCompilation evaluates the existing validators once and constructs HtmlValidationOutcome in the producer-owned BlockCompilationOutput.
• DiagnosticsCollector consumes the compact outcome rather than full block-validity, semantic-parity and content-round-trip reports.
• Keep only original finding keys used by diagnostics, preserving mixed values and leaving existing fallback messages/defaults with the collector.
• Uncomputed outcomes explicitly say not_evaluated, not pass. Successfully parsed empty documents still evaluate validators normally.
• Preserve full reports and conversion-report projections, canonical serialization, status/acceptance policy and all block/artifact capabilities.

There is no reduced-evidence mode, skipped validator or performance claim in this bite. Detailed reports are still computed; this establishes the consumer boundary needed before a later evidence policy. The internal outcome is not added to the public serialized result.

The required/detail consumer inventory is documented in php-transformer/docs/html-validation-outcome-consumer-inventory.md.

Verification

Baseline: c302eaaa26681462a784c0d341f2702ce7d464e7.
Candidate: b36079ffb337b3518a93fc3d6b8a54cb27907ac8.

• All 385 discovered HTML fixtures retain identical serialized-block hashes and complete result envelopes after excluding only recursive transform_duration_ms. PHP serialization preserves binary bytes and array ordering.
• Nine targeted HTML/artifact cases also match, including nonzero companion, runtime, stylesheet, shared-shell and responsive-counterpart outputs.
• Coordinator independently compared the complete baseline/candidate evidence files; both have SHA-256 a3a2c5c327cace26a2a69aaa52d9e81fbf7e2fef9134259678700c8517d28d56.
• Direct diagnostic assertions cover failure findings for all three validators, missing/null summary fallbacks, mixed severity/location values and uncomputed outcomes.
• Full composer test passed end-to-end in one PHP 8.4 container with a 512 MB PHP configuration inherited by child processes: canonical/unit tests, 303 parity fixtures, dist shape and package-install proof.
• Baseline and candidate used the same task-local Composer lock for matching dependencies; the repository does not commit a lock.
• WordPress integration was skipped locally because WP_TESTS_DIR was unavailable. Actual CI integration and solved-site gates remain required.

Reproduce

From php-transformer, with dependencies installed and PHP memory configured for child processes:

composer test
REQUIRE_WP_TESTS=1 WP_TESTS_DIR=/path/to/wordpress-tests composer test:wordpress-integration

For corpus comparison, run this from both repository roots with matching dependencies and compare the resulting output:

php -d memory_limit=512M -r '
require "php-transformer/vendor/autoload.php";
$strip = function (array &$value) use (&$strip): void {
    unset($value["transform_duration_ms"]);
    foreach ($value as &$item) if (is_array($item)) $strip($item);
};
$rows = [];
foreach (new RecursiveIteratorIterator(new RecursiveDirectoryIterator("fixtures", FilesystemIterator::SKIP_DOTS)) as $file) {
    if (!$file->isFile() || strtolower($file->getExtension()) !== "html") continue;
    $result = (new Automattic\BlocksEngine\PhpTransformer\HtmlToBlocks\HtmlTransformer())->transform(file_get_contents($file->getPathname()));
    $array = $result->toArray();
    $strip($array);
    $rows[$file->getPathname()] = [hash("sha256", $result->serializedBlocks), hash("sha256", serialize($array))];
}
ksort($rows, SORT_STRING);
echo json_encode($rows, JSON_UNESCAPED_SLASHES), PHP_EOL;
'

AI Assistance

OpenAI GPT-5.6 Terra via direct OpenCode mapped consumers, implemented the boundary and ran lab verification. GPT-6 Astra via OpenCode coordinated and reviewed the work, identified unevaluated-status and diagnostic-normalization problems, requested behavioral corrections and full-suite verification, independently checked comparison evidence, and prepared this PR under Chris Huber's direction. No release or deployment was performed.