Admin page for catalogue AI-translation freshness (block-6 plan, Task 5): /admin/catalogue/translations.

Lists (resource, target language) rows across the four catalogue translatable types (item / category / set_label / set_value), filterable by type, language, freshness state and a name search, with a per-state count in the header. Row actions enqueue a single translation (Translate) or record the current source as the reference without calling the AI (Stamp fresh, TranslationStatus.stamp_fresh/2); bulk actions do the same across every row matching the current type/language/ search scope for one target state ("Translate all missing" / "Retranslate all stale").

Only rendered when PhoenixKitAI is configured with a working endpoint — otherwise a notice points at the AI admin section. Availability and the endpoint/prompt resolution are the SAME check the sweep tick makes (TranslationSweepWorker.endpoint_and_prompts/0), so a manual click here and an automatic sweep enqueue can never disagree about which endpoint or prompt a resource type uses.

In-flight tracking

PhoenixKitAI.Translations.job_in_flight?/1 (the private check enqueue/1 itself uses for dedup) isn't part of the module's public API, so rather than re-querying oban_jobs directly this page tracks in-flight rows the same way phoenix_kit_ai's own single-resource AITranslate component does: an enqueue/1 call's conflict?/:ok result marks a row, and the global {:ai_translation, event, payload} broadcast (Translations.subscribe/0) clears it again on completion/failure — no private function, no direct DB query.

Gettext

Uses the local gettext/1,2 macros (use Gettext, backend: PhoenixKitCatalogue.Gettext), not the qualified Gettext.gettext(Backend, ...) runtime call some older catalogue modules still carry — the qualified form is a plain function call the extractor never sees (see commit "Fix strings the gettext extractor cannot see").