build(i18n): give each language a flat file so a translation diff is reviewable

[datlechin]

Localizable.xcstrings interleaves every language inside every key, so a translator changing fifty Vietnamese strings edits a 109,196-line file next to Turkish and Chinese they never touched. That is the real reason translation contributions stall, and the app ships four non-English languages sitting at 3,827 of 3,829 strings each with no route for anyone outside the project to move them.

scripts/localization.py gives one flat file per language:

scripts/localization.py status              # coverage per language
scripts/localization.py export vi           # Localization/mac.vi.json
scripts/localization.py import vi           # merge back
scripts/localization.py verify              # round-trip is byte exact

--target ios covers the iPhone catalog, which is separate.

Why this is not Crowdin

Crowdin was the obvious answer and I could not verify it works here. Its Apple String Catalog support is an exporter: the store page describes it as generating .xcstrings files from Crowdin project strings, which is the opposite direction from ingesting one as a source file. I could not confirm .xcstrings is accepted as a source format before this session's search budget ran out, so shipping a crowdin.yml would have been a guess. Weblate is already ruled out: it cannot read String Catalogs at all.

This needs no service, no account and no decision, and it does not block adopting a platform later.

Why the exported file is not committed

Because it turned out not to be necessary. The 109k-line problem is caused by reformatting the catalog, not by the format itself. Once a merge is byte exact for everything it did not translate, one changed string produces a two-line diff in the catalog, which is perfectly reviewable. Verified: changing a single Vietnamese string gives 1 file changed, 1 insertion(+), 1 deletion(-).

So Localization/ is ignored, the catalog stays the single source of truth, and the repository does not gain four megabytes of duplicated strings that can drift.

Byte-exactness needs three Xcode quirks reproduced, all covered by verify: the key separator is " : " rather than ": ", an empty object is written across three lines rather than as {}, and there is no trailing newline. Keys keep the catalog's own order; sorting them looks tidier and rewrites the whole file.

Two bugs the round-trip test caught

Both were mine, and both would have corrupted translations silently.

.strip() on the merged value destroyed meaningful whitespace. It was meant to detect a blank translation, but it also rewrote the stored string: "\n\n… (%d characters not shown)" came back as "… (%d characters not shown)" in all four languages. Blankness is now tested against the stripped form while the raw value is stored.

Re-importing an unedited file flipped needs_review to translated on 14 strings. State belongs to whoever reviewed the string, not to whoever last ran the script, and rewriting it also meant a no-op round-trip produced a diff, defeating the point. The merge now writes state only when the value actually changed, and the export surfaces state read-only so a translator can find what still needs attention.

Verified on all six language/catalog pairs: export then import with no edits leaves both catalogs byte identical.

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.