[kasnder]

Refs #1009 (the translation half; the crash report is still pending).

Summary

1. Broken translations (490c221). The bulk fill in #940 left filler words in place of real text in many locales: "Elemento", "элемент", "عنصر", "Atriði", "asia" and others, plus stub phrases such as "Ulteriori informazioni su …". Every translated string in every locale was checked against the English, and 2,111 strings that were actually wrong were replaced:

• The worst files were Icelandic (339), Galician (311), Arabic (282), Basque (256), Finnish (245), Russian (229) and Italian (190).
• Locales without filler still had real errors:
  • "start" translated as "stop" in both Chinese locales;
  • Private DNS "enabled" where the English says "disabled" in Slovenian;
  • opposite meanings in Dutch and French;
  • a full-width ％d that broke the Japanese count;
  • translated WireGuard config keywords ([Interface] and others);
  • wrong GDPR abbreviations in Polish and French.
• Correct but imperfect translations were left alone.

2. Missing strings and plurals (9a634ca).

• Missing strings: each locale lacked 5–58 strings that exist in English, about 1,370 in total. They are now translated, using the terms and form of address each locale already uses.
• Plurals: every plural resource now has exactly the categories its language needs (CLDR):
  • Before this PR all locales had only one/other, so Russian, Ukrainian, Polish and Czech showed the wrong word form for counts like 2–4 and 5–20.
  • Arabic, Hebrew, Slovenian and Romanian gain the categories they were missing.
  • French, Spanish, Italian and Portuguese gain many.
  • Japanese, Korean, Chinese and Vietnamese keep only other.

How

• Sonnet agents reviewed and translated the locales and wrote their proposed changes as JSON.
• A script then applied them to the XML in place and rejected any change whose format specifiers, HTML tags, CDATA wrapper, \n count, quote escaping or plural categories didn't match the English. Nothing failed those checks.
• One leftover filler string the review missed (Galician msg_started) was fixed by hand.

Testing

• All 36 strings.xml files parse as XML.
• No filler phrases are left in any locale.
• Every locale now has every translatable English resource.
• Not run here: the container has no Android SDK, so CI is the first build and lint check.

Before and after merging

• Crowdin: the daily Crowdin sync downloads translations and opens a PR, which could put the old strings back. After merging, run Synchronize Crowdin translations by hand with upload_translations ticked, then check that Crowdin took the new versions.
• Not reviewed by native speakers: these are machine translations. The long rewrites, especially the GDPR request and erasure letters, are worth a native speaker's look. The agents flagged 39 entries they weren't sure about and left them unchanged.

https://claude.ai/code/session_01Ggf8mGDTZKcAduhWquDHAj

---

Generated by Claude Code