Localization: AI translation primitives

[jkmassel]

Reusable Ruby primitives for the AI translation tier of the localization pipeline — the service behind the human ?? AI ?? English floor whose AI stub (ai_translate_plural → nil) was left open in #25688. All of it is pure prompt-building + validation with the Anthropic SDK call injected, so every line of logic is unit-testable without the gem or the network; the live SDK wiring is one thin factory.

Nothing is wired into a lane yet — these are the building blocks, deliberately decoupled from the GlotPress / catalog plumbing that's still in flux.

Summary

• TranslationValidator — the format-specifier safety gate. A machine translation must preserve the source's printf/NSString arguments exactly (count + type; positional %1$@ may reorder, which is the whole point). A mismatch is rejected, so a broken translation falls back to English rather than shipping a crash in a locale no one on the team can read.
• Glossary — brand do-not-translate list (WordPress, Jetpack, …) plus per-locale preferred terms and a register note, rendered into the prompt. Pure data in; sourcing it (the WordPress.org per-locale glossaries / style guides) is pre-processing handed in later.
• AITranslator — three translation shapes: translate (one string), translate_plural (a whole CLDR form-set in one request, so the model keeps one stem across forms), and translate_all (batched regular strings). Structured outputs (output_config json_schema) enforce the reply shape on the plural and batch paths.
• AnthropicBatch — the async Message Batches path for a bulk backfill (~50% cheaper): submit → await (poll) → results → collect_batch, plus all the SDK-shape glue, shared with the sync path so the request shape can't drift between them.

Design notes

The SDK call is injected

AITranslator takes a complete: callable (and the batch path takes a client), so prompt-building, validation, batching, and result-assembly are all exercised by unit tests with a canned reply. AITranslator.with_anthropic / AnthropicBatch.client build the live instances (default model claude-opus-4-8). The two live-API bugs we hit — a custom_id that didn't match ^[a-zA-Z0-9_-]{1,64}$, and results_streaming yielding raw JSONL strings rather than typed objects — were both invisible to a permissive fake client and only caught by real calls. The SDK seams are now live-verified, not just fake-tested.

The placeholder gate is a hard floor

Every machine cell — single string, plural form, or batch entry — passes through TranslationValidator before it's returned. This is the same invariant the catalog needs_review machinery already assumes: the AI tier can only ever produce a safe translation or nil.

Plural consistency

Translating each CLDR category independently let the model drift between synonyms across forms (Polish słowo → wyrazy → słów). translate_plural sends the whole form-set in one request and instructs one consistent stem; verified it now yields słowo / słowa / słów.

Verified live

Against the real API (fr/de/ja/pl): verb/adjective disambiguated from the dev comment (Suivre/Suivi, Folgen/Gefolgt), brand terms kept verbatim, German informal register (Dein), French space-before-?, plural stems consistent, and a full Batch round-trip (submit → await → collect) returning Suivre / %1$@ vues mapped back to keys and grouped by locale. Placeholders were preserved throughout — the gate never had to fire on real output.

Not in this PR (deliberate)

• Lane wiring — download_localized_plurals still calls the nil stub; switching it to AITranslator#for_plural (and a translate_all pass over the regular catalog) is a separate change.
• A .strings / .xcstrings reader — these consume a {key, source, comment} array and plural form-sets; building that from the real catalog (plus any pre-processing) is upstream of here.
• Glossary sourcing — Glossary is data-in; pulling the WordPress.org per-locale glossaries / style guides into it comes later.
• Bulk-backfill ergonomics — model tiering for cost, and the submit-now / collect-later split for very large batches (vs. the synchronous await).

Test plan

All checks are pure Ruby — stdlib minitest, no bundle, no network:

• ruby fastlane/lanes/translation_validator_test.rb — 9
• ruby fastlane/lanes/translation_glossary_test.rb — 5
• ruby fastlane/lanes/anthropic_batch_test.rb — 6
• ruby fastlane/lanes/ai_translator_test.rb — 30
• rubocop clean on all eight files
• Live translation pass (needs ANTHROPIC_API_KEY + bundle install): ruby fastlane/lanes/ai_translator.rb fr "You have %1$d new posts" "Notification. %1$d is the count."

Related

• Localization: String Catalog pipeline (plurals + catalog generation) #25688 — String Catalog pipeline; this fills its AI-tier stub.

[dangermattic]

	1 Warning
⚠️	This PR is larger than 500 lines of changes. Please consider splitting it into smaller PRs for easier and faster reviews.

	1 Message
📖	This PR is still a Draft: some checks will be skipped.

Generated by 🚫 Danger

[wpmobilebot]

📲 You can test the changes from this Pull Request in Jetpack by scanning the QR code below to install the corresponding build.

	App Name	Jetpack
	Configuration	Release-Alpha
	Build Number	32867
	Version	PR #25705
	Bundle ID	com.jetpack.alpha
	Commit	7412850
	Installation URL	2brteotvt3r8g

Automatticians: You can use our internal self-serve MC tool to give yourself access to those builds if needed.

[wpmobilebot]

📲 You can test the changes from this Pull Request in WordPress by scanning the QR code below to install the corresponding build.

	App Name	WordPress
	Configuration	Release-Alpha
	Build Number	32867
	Version	PR #25705
	Bundle ID	org.wordpress.alpha
	Commit	7412850
	Installation URL	18r3gimi0qglo

Automatticians: You can use our internal self-serve MC tool to give yourself access to those builds if needed.