# `i18n`
Five-locale internationalisation — the Rust half of `@evinvest/i18n`.
```toml
ev_lib = { version = "0.13", default-features = false, features = ["i18n"] }
```
`wasm32`-safe, no I/O, and zero dependencies *there*. Natively it carries one,
`inventory` — see [Extraction is linking](#extraction-is-linking).
## Why a Rust half exists
`real_estate_allocation` is a Dioxus app. It cannot import an npm package, and a
zone rendering English chrome inside a Russian shell is exactly the failure the
shared locale registry was created to prevent.
The two halves are not merely similar — they read **the same
`messages/<locale>/*.json` files** and apply the same rules, so a catalogue
authored for the site works unchanged in a zone and neither half can drift
alone. `tests.rs` pins that: it renders the site's shipped plural patterns and
asserts the output matches what `Intl.PluralRules` + `Intl.NumberFormat` produce
in the browser.
## What's here
| [`Locale`] | the five locales, their codes and their native labels |
| [`locale_path`] / [`split_locale_path`] | the `/<locale>` URL contract — default unprefixed, everything else prefixed |
| [`negotiate`] | `Accept-Language`, with q-values and regional-tag matching |
| [`t!`] / [`Translator`] / [`format_message`] | ICU-subset formatter: `{name}`, `plural`, `select`, `#`, `'` quoting |
| [`catalogue`] | every `t!` site linked into this binary, as `messages/en/common.json` |
| [`policy`] | rules 1.1–1.3: English is canonical, stale translations are refused, untranslated content is hidden |
## English is written where it renders
```rust
t!(tr, "portfolio.why.title", "Why Quy Nhon?")
```
The key *and* the English sentence, at the call site;
`messages/en/common.json` is generated back out of the code. A key a translated
catalogue lacks renders the sentence the component asked for.
This reverses what this module used to document. Rendering the dotted key was
defensible while the catalogue was the only place the English lived — it was
greppable and survived to a screenshot. Under the new contract the call site has
the correct sentence in hand, so a dotted key on screen is strictly worse.
`Translator::has` went with it: its only purpose was branching around that
placeholder.
`$key` and `$en` are `literal` fragments, so **the compiler is the literal-ness
gate** — the TypeScript extractor has to enforce that by hand.
## Extraction is linking
`t!` also registers its pair, so [`catalogue`] returns every site linked into the
binary — whether or not it ever renders. The alternative, a recording
`Translator` driven by one SSR pass, would silently drop every string behind a
branch that did not run, which is the exact failure class this is here to catch.
That is what `inventory` buys, and it is native-only: a shipped wasm bundle
extracts nothing and stays zero-dep.
**The ceiling: copy must live in natively-compilable code.** For a microfrontend
the component-MFE snapshot contract already requires exactly that — the same
`view` components render the live bundle and the static snapshot — so the
constraint costs a producer nothing it was not already paying.
Wire the two halves of the TypeScript `evinvest-i18n-{extract,check}` pair as
one example and one test:
```rust
// examples/dump_messages.rs — writes messages/en/common.json
// tests/catalogue.rs — asserts the committed file still matches, and
// runs policy::audit over the translated locales
```
## The two things worth knowing before you change this
**Plural rules are hand-written, and that ceiling is deliberate.** `Intl` has no
`std` equivalent, and ICU4X would put megabytes of locale data into a wasm
bundle. The rules in `plural.rs` are transcribed from CLDR 46 for exactly five
locales. Adding a sixth means transcribing its rule by hand — which is the
review moment a generic fallback would quietly rob us of. `plural_categories`
and `select_plural` must agree; a test asserts it, because the policy's plural
check is only sound if they do.
**`other` is not a wildcard in the policy.** The formatter falls back to `other`
at runtime so a view cannot break — that is not evidence a translation is right.
A Russian `one`/`other` pair renders "5 участка" for five, which is grammatical
nonsense the runtime cannot detect, *because it got a branch*. The policy
therefore requires every category the locale declares. Vietnamese declares only
`other` and so passes on the same rule, not an exemption.
## Formatting numbers and money is not this module's job
`plural` and `select`, yes. `number`, `date`, `currency`, no. A consuming app
owns one policy per unit of measure and interpolates the already-formatted
string; a second, competing number policy hiding inside message catalogues is
precisely the drift that rule exists to prevent. The one exception is `#` inside
a plural branch, which *is* the count — see [`Locale::format_number`].
[`t!`]: https://docs.rs/ev_lib/latest/ev_lib/macro.t.html
[`catalogue`]: https://docs.rs/ev_lib/latest/ev_lib/i18n/fn.catalogue.html
[`Locale`]: https://docs.rs/ev_lib/latest/ev_lib/i18n/enum.Locale.html
[`Locale::format_number`]: https://docs.rs/ev_lib/latest/ev_lib/i18n/enum.Locale.html#method.format_number
[`Translator`]: https://docs.rs/ev_lib/latest/ev_lib/i18n/struct.Translator.html
[`format_message`]: https://docs.rs/ev_lib/latest/ev_lib/i18n/fn.format_message.html
[`locale_path`]: https://docs.rs/ev_lib/latest/ev_lib/i18n/fn.locale_path.html
[`split_locale_path`]: https://docs.rs/ev_lib/latest/ev_lib/i18n/fn.split_locale_path.html
[`negotiate`]: https://docs.rs/ev_lib/latest/ev_lib/i18n/fn.negotiate.html
[`policy`]: https://docs.rs/ev_lib/latest/ev_lib/i18n/policy/index.html