Avoid blindly re-encoding HTML files

Previously, HTML files werei stripped of their XML
Processing Instruction headers and re-encoded from
UTF-8 to HTML-ENTITIIES to be fed into the DomDocument.

This caused problems for documents with CDATA blocks that
contained Unicode, as it's not correct to escape that as
HTML entities in the general case. For example CSS or
binary data doesn't use that escaping system.

Instead, load it directly and then remove the PI nodes
after the fact.

Bug: https://phabricator.wikimedia.org/T271390

src/Util/Util.php

@@ -185,10 +185,26 @@ public static function extractErrorMessage( ?ResponseInterface $resp, RequestInt
 	 */
 	public static function buildDOMDocumentFromHtml( string $html ): DOMDocument {
 		$document = new DOMDocument( '1.0', 'UTF-8' );
+
 		libxml_use_internal_errors( true );
-		$document->loadHTML( mb_convert_encoding( str_replace( '<?xml version="1.0" encoding="UTF-8" ?>', '', $html ), 'HTML-ENTITIES', 'UTF-8' ) );
+		// enforce a UTF-8 encoding declaration
+		$document->loadHTML( '<?xml encoding="utf-8" ?>' . $html );
 		libxml_clear_errors();
+
 		$document->encoding = 'UTF-8';
+
+		// Dirty fix to strip out existing XML Processing Instruction nodes
+		// (we already have one from the creation of the DOMDocument)
+		//
+		// This is better than re-encoding from UTF-8 to HTML-ENTITIES, because
+		// that will escape things even in CDATA blocks (e.g. T271390)
+		// https://www.php.net/manual/en/domdocument.loadhtml.php#95251
+		foreach ( $document->childNodes as $item ) {
+			if ( $item->nodeType === XML_PI_NODE ) {
+				$document->removeChild( $item );
+			}
+		}
+
 		return $document;
 	}