mirror of
https://github.com/bookorbit/bookorbit.git
synced 2026-11-03 00:11:37 -06:00
neon
3b9bae7525
Two independent guards were failing the workflow on healthy input. Retention read "exported value equals the English source" as a lost translation, but exports are built with skipUntranslatedStrings, so Crowdin omits a key entirely once it stops carrying a translation. Anything present in the export is by definition translated, and the equality check only ever fired on correct work: Spanish "Error", "Feb" and "Cookie", French "Description", Portuguese "Readwise". Retention now compares presence alone, which is what absence already meant. The same inference lived in a legacyComplete branch that would have silently stopped protecting the 2,318 keys whose translation legitimately equals the source once any locale reached full translation. Scheduled runs had no way out of that. The only escape hatch is allowed_translation_losses, a workflow_dispatch input that is never persisted, so every schedule rediscovered the same losses and failed until someone dispatched the run by hand. The slot count rule then only runs for the keys it is given, and the sync passed none, so its rejection filter was weaker than the locale validation running on its output one step later. An invalid ja message survived the filter, reached the catalog, and failed the next step instead of being dropped and reported. Move the source scan into locale-source-keys.mjs so the sync and validate-locales.mjs derive the same keys.