i18n Localization

Portability: 3 stdlib Python scripts. No network, no translation service. They check and generate files; they never translate text or call an API.

A release with a new language usually breaks in the same places: keys added in English and never synced, a placeholder renamed by a translator, Russian or Arabic files that copied English's two plural forms, UI text that never went through t(), and dates formatted in the server's locale. This skill makes each of those a check that fails before the release, not a bug report after it.

Workflow

# 1 - Translation files: keys, placeholders, ICU syntax, CLDR plural forms
python scripts/locale_sync.py locales/ --source-locale en
# 2 - Code: user-facing text that bypasses t(), concatenation, locale-less formatting
python scripts/hardcoded_strings.py src/
# 3 - Layout: generate a pseudo-locale and click through the app in it
python scripts/pseudo_localize.py locales/en/common.json -o locales/en-XA/common.json
python scripts/pseudo_localize.py locales/en/common.json --bidi -o locales/ar-XB/common.json

Step 1 reads JSON/ARB (i18next or ICU), .properties, Apple .strings, Android strings.xml and gettext .po, pairs each target with the source file of the same namespace, and blocks on missing or empty keys, changed placeholders, ICU syntax errors, plural/select without other, a zero branch that can never match, and plural categories the target locale needs (ru few/many, ar six forms), using the embedded CLDR 49 table. Untranslated values, orphan keys, fuzzy entries and markup drift are warnings.

Step 2 blocks on literal text between tags and literal user-facing attributes (placeholder, alt, aria-label, title...), and warns on literals rendered from {cond ? 'Saving...' : ...}, t('a') + x concatenation, toLocaleString() with no locale, '$' + amount and hardcoded date patterns. It is a text heuristic: read each finding. Use --warn-only to adopt it on a legacy codebase, i18n-ignore on a line, and .i18n-allow for brand names.

Step 3 accents, pads (by translation-growth rates) and brackets every string while keeping placeholders and ICU syntax intact. Plain English on screen is a hardcoded string; a cut-off [[... is truncation; [[a]] [[b]] is concatenation.

Hard rules

Never concatenate translated fragments. One message per sentence, with named placeholders and a plural branch.
Plural forms follow CLDR for the target locale, not English. Never copy _one/_other into ru, pl, uk, cs, ar, he files and call them done.
=0, not zero, for an empty-count message in ICU.
Pass the locale to every formatter. Currency comes from the data, not the UI locale.
Placeholders are code. A translator may move them, never rename or drop them; use positional %1$s where order can change.
Do not machine-translate to make the checker pass. A missing translation falls back to the source language; a wrong one ships silently. Mark gaps for a translator instead.

Scripts

Script	Role	Exit codes
`scripts/locale_sync.py`	Translation files vs source locale: keys, placeholders, ICU, CLDR plurals, gettext	0 · 2 warnings · 3 bad input · 4 blocked
`scripts/hardcoded_strings.py`	React/Vue/Svelte/HTML: untranslated text, concatenation, locale-less formatting	0 · 2 · 3 · 4
`scripts/pseudo_localize.py`	`en-XA` (or `--bidi` `ar-XB`) pseudo-locale from JSON/ARB/.properties	0 · 3

All support --help, --sample and --output json.

References and assets

references/plurals_and_messages.md: CLDR categories per locale, ICU rules, library key shapes, placeholders (8 sources)
references/string_routing_and_formatting.md: what to route through t(), Intl formatting, BCP 47 and fallback (7)
references/layout_rtl_pseudo.md: text expansion, pseudo-locales, RTL, release checklist (7)
assets/i18n-gate.yml: CI job example
assets/i18n-allow.example: allowlist format