feat(i18n): translation foundation, isolated from the host app's I18n

[AnjanJ]

RED had no i18n at all — no config/locales, no t() call anywhere. This adds the machinery to translate the dashboard without translating anything yet. Every string is still English, and rendered HTML is byte-identical to main.

The #148 fix added config.dashboard_locale and an around_action to stop Pagy's locale leaking, and described itself as groundwork "so the mechanism is in place for translating the rest of the dashboard." This is that mechanism.

Why a private I18n backend

RED translates through its own I18n::Backend::Simple rather than the host's. Sharing the host's backend would hand it three ways to break the error dashboard — the page that has to work when everything else is broken:

• raise_on_missing_translations turns a key we forgot into a 500
• enforce_available_locales raises as soon as RED asks for its own locale
• a custom exception_handler can raise on anything

For the same reason the around_action does not call I18n.with_locale, which mutates host global state. RED's locale lives on its own CurrentAttributes and is cleared in the ensure, so it cannot outlive the request or strand on a recycled Puma thread.

Pagy's locale and RED's resolve independently — they ship different dictionaries, and a locale RED can serve but Pagy cannot must still render, with English pagination.

The helper is red_t, not t: overriding t risks colliding with host helpers and with Rails' lazy lookup, and an explicit name keeps every translated site greppable when the remaining ~1,600 strings get extracted.

Trade-off accepted: hosts cannot override RED's strings with their own locale files. That is the right default for a self-hosted ops tool, and can be relaxed later without breaking anything.

Two bugs only found by running it

Both would have been 500s on the error dashboard:

1. A raw backend lookup throws :exception on a miss rather than returning — I18n.translate is what normally catches it. Uncaught, every missing key becomes an UncaughtThrowError.
2. I18n::InvalidPluralizationData raises when a locale supplies only other and the count needs one. English has both forms, so this only bites once a real translation ships.

Both are caught and degraded to English. Nothing in I18nStore can raise.

Dates

Date formats are now translation keys rather than literals — "%B %d, %Y" is a US ordering as much as it is English words. local_time_ago interpolates instead of concatenating " ago", which is the wrong word order in several languages.

Known limitation: local_time_ago's server-rendered fallback still calls Rails' time_ago_in_words, which translates through the host's I18n. The browser replaces that text on load, so it only shows for non-JS clients. Phase 3 localizes the JS side.

Scope

Full UI translation is deferred to v1.1+ (ROADMAP item AA). It is ~1,600 strings across 48 views, inline JS, mailers, and notification payloads — larger than any feature shipped so far — and it is reach rather than depth, which the roadmap already defers on the same grounds as APM.

Shipping the foundation now makes config.dashboard_locale honest (it previously promised more than it delivered) and lets pages be translated one release at a time instead of in a single push. No user has asked for this; the demand signal is zero i18n issues filed.

Test plan

• RSpec: 3743 examples, 0 failures (+95), green on a second seed
• RuboCop: 494 files, no offenses
• Chaos tests: 1442/1442 assertions across all 5 apps, production mode
• Rendered HTML diffed against main across 4 pages — identical after normalizing CSRF tokens, timestamps, and Faker data
• Mutation-tested: deleting the ensure cleanup fails 3 specs; deleting the English fallback fails 1

An earlier version of the leak specs passed even with the cleanup removed — CurrentAttributes resets between requests in the test harness, so asserting Current.locale.nil? after a get tested a framework guarantee rather than the code. Found by deliberately reverting the production change and watching them stay green. They now exercise the around_action directly.

• Verified the built gem contains config/locales/en.yml — no gemspec change needed, it already ships config/**/*

Refs #148

🤖 Generated with Claude Code