feat(app): i18n strategy — app-only en+ru with a synced translation pipeline (CLEAN-33)

[maksymhryzodub-prog]

Implements CLEAN-33: https://dreamvention.atlassian.net/browse/CLEAN-33

The ticket read as "the i18n module is wired up, only the locale list is missing". The code said otherwise: across both consoles only 5 of 307 .vue files call t(), the locale files hold 70 keys total, six of them are empty {}, and no console renders a switcher. The real work is extracting hardcoded strings, not translating them — so this starts with a decision about scope.

Full reasoning: docs/superpowers/specs/2026-08-19-i18n-strategy-design.md.

Decisions

• Only app/ gets localized. admin/ stays English by choice, not by deferral — it is an internal operator tool, and localizing it would mean touching ~270 files to serve people who read English anyway.
• en (default) + ru, no_prefix strategy — app is a SPA, so /ru/ routes would add surface with no SEO payoff. detectBrowserLanguage already handles the first render; the new switcher lets a user override it.
• One LOCALES constant every slice registers, instead of repeating the locale list per slice.
• Translations are generated locally and verified in CI, rather than autotranslated by a bot in the PR — no LLM key in repository secrets, no bot commits, and the check still makes it impossible to forget.
• Out of scope, deliberately: admin/, the language the agent answers in (that's a prompt/agent setting, not UI i18n), and API error text.

What's in here

• scripts/i18n-sync.ts — diffs each slice's en.json against ru.json and sends only missing or stale keys to claude-opus-5, one request per slice, writing results back in en.json's key order.
• app/i18n.sync.json — hash of the English value at translation time. Without it only new keys would ever be noticed and an edited English string would keep its old translation forever.
• Adopt rule: a key the manifest never recorded is adopted, not retranslated — a hand-written or hand-corrected translation is never overwritten.
• bun run i18n:check in CI — key sets and hashes only, no network and no API key, so it runs on forks too.
• Locale switcher in the app shell + fallbackLocale, so a not-yet-extracted key renders English instead of a raw key path.
• Extraction: bridle (2 strings were left) and common in full — shell, landing hero, demo card, landing page. common came along because the switcher lives there and a half-translated first screen reads worse than an English one.

Verification

• nuxt typecheck and nuxt build clean; the Russian strings appear in the built client chunks.
• i18n:check exercised against all three drift modes — new key, edited English source, key removed from en.json — each exits non-zero with the offending keys listed.
• The translate path ran for real: ru.json and the manifest were deleted and rebuilt by the script (71 keys). A second run reports nothing to do, and a hand-correction on top of the output survives the next sync.

Follow-up

Remaining slices, one PR each, in the order a user meets them: agent → chat → user → template. chat and user have no i18n block yet; they get one when their turn comes.

🤖 Generated with Claude Code