Harden mojibake repair with Czech-specific serialized value handling

2026-09-08 15:09:13 +02:00
parent 4c623c36ea
commit 25de2a13f9
3 changed files with 413 additions and 82 deletions
@@ -0,0 +1,56 @@
+# Mojibake Repair (Temporary)
+
+Version 1.1.0 replaces the original whole-column SQL conversion with targeted
+replacements for Czech letters misdecoded as ISO-8859-1 or Windows-1252.
+
+## Why version 1.0.0 could break fonts
+
+The original conversion also changed `options` and metadata. Theme and builder
+settings can contain PHP-serialized strings with embedded **byte lengths**.
+Changing the text without updating those lengths makes the settings unreadable;
+a theme can then fall back to default typography. Converting entire values could
+also damage already-correct Unicode in mixed content. WordPress documents this
+serialization risk in its [migration guide](https://developer.wordpress.org/advanced-administration/upgrade/migrating/).
+
+## Sites already repaired with version 1.0.0
+
+1. Deactivate version 1.0.0 and save a copy of the current database.
+2. Use a pre-repair backup to compare and recover affected theme/builder settings,
+   preferably on staging first. A full database restore can discard newer content
+   and orders, so choose the recovery scope using the site's actual changes.
+3. Run the new version's preview against the restored data and review the changes.
+   The new version prevents this serialization damage; it does not reconstruct
+   settings already corrupted by the previous conversion.
+4. After repair, clear page caches and regenerate builder CSS if the site uses it.
+   Verify fonts and content on the site, then deactivate and remove the plugin.
+
+## Behavior and limits
+
+- Repairs recognized Czech mojibake sequences without transcoding whole values.
+  Other corruption patterns and JSON `\uXXXX` escapes are outside this repair's
+  scope. A sequence intentionally shown as a mojibake example will also match,
+  so review the preview.
+- Preserves CSS, JSON escaping, serialized keys and scalar formatting. Repairs
+  serialized array/string values, including nested serialized strings, and
+  recalculates their byte lengths without instantiating PHP objects.
+- Leaves an entire candidate value unchanged if it contains malformed or
+  unsupported serialization, objects, references, excessive nesting, or invalid
+  UTF-8. The preview reports skipped counts and example row IDs. This is not a
+  general audit of already-corrupted settings: values without a recognized
+  mojibake sequence are left untouched without further inspection.
+- Reads in batches of 200. Execution uses an InnoDB transaction and checks the
+  exact candidate IDs and values against the signed preview. Changes since
+  preview or write conflicts cause rollback. Old-version approvals are rejected.
+- Clears the WordPress object cache after commit. External page caches and
+  generated builder styles may still need to be refreshed separately.
+- Scans the current site's fixed core-table plan. Large databases may exceed
+  hosting request limits; batching bounds the rows loaded at once, not the total
+  request duration or transaction size.
+
+## Local verification
+
+Run `php -l wp-mojibake-repair.php` and `php tests/repair.php`.
+The regression script requires PDO SQLite and iconv. It checks Czech text,
+serialized typography settings, preserved CSS/JSON, skipped unsafe values,
+batching, preview approvals, rollback, and cache invalidation. Database-flow
+checks use a SQLite adapter; they do not replace verification on WordPress/MySQL.