[Translation] Make CrowdinProvider::write() add the locales missing from the project - #66028

Rebased on 8.2 and amended before merge.

The locale to language id conversion was str_replace("_", "-", $locale). Of the 314 languages Crowdin lists, only es-ES, la-LA and pt-PT use that shape, so fr_FR, de_DE, it_IT, ja_JP and most European locales produced an id that does not exist and the patch came back with Language fr-FR not found. write() now reads GET /languages once, only when a locale is missing, and maps a locale to an id through the locale field each language carries, preferring an exact id match for the three ambiguous cases.

Unsupported locales are now left out of the patch instead of travelling with the others, so a single unknown locale no longer stops every other one from being added; they keep falling back to the Ignored "%s" locale ... warning. A locale whose language the project already has, such as fr_FR against fr, no longer triggers a patch at all.

A 5xx on either request now throws like the rest of the bridge, while a 403 still only logs and skips. Added an UPGRADE-8.2.md note about the project.settings write scope, plus four tests.

I left out an RFC 6902 test op guarding the read-modify-write window: test on an array compares member by member including order, nothing promises Crowdin returns targetLanguageIds in a stable order, and a Crowdin that does not honour it would turn every push into a 400. The window is one round trip.

Thank you for the rebase and the amend, and above all for catching the language id mapping. I had assumed str_replace('_', '-', $locale) was enough, and reading GET /languages once is clearly the right call when only 3 of the 314 languages use that shape. Leaving unsupported locales out of the patch instead of letting one of them sink the others is a real improvement too.

On the current red, Psalm fails on the new getSupportedLanguageIds():

CrowdinProvider.php:563 InvalidReturnType: declared 'array<string, string>', got 'array<array<array-key, string>|string, mixed>'

$language['data']['locale'] comes out of toArray() as mixed, so Psalm allows str_replace() to return an array and never narrows the key to string. Casting before use should settle it:

foreach ($response->toArray()['data'] as $language) {
    $id = (string) $language['data']['id'];
    $languageIds[str_replace('-', '_', (string) $language['data']['locale'])] ??= $id;
    $languageIds[str_replace('-', '_', $id)] = $id;
}

Want me to push that, or would you rather fold it into your amend?

The two unit test reds look unrelated: the same testSchemeRedirect and DoctrineOutboxIntegrationTest failures show on #65989, and both were fixed on 8.2 right after this run (6cc2cdda and b96ffaca). A rebase should clear them.

Thanks, and you are right on both counts. The Psalm error is mine, so I folded the casts into the amend rather than have you push a fix for it: toArray() hands back mixed, so str_replace() could return an array as far as Psalm knows, and the key never narrows to string. Taking $id once as a (string) also reads better than casting it twice.

Rebased on 8.2 as well, so the testSchemeRedirect and DoctrineOutboxIntegrationTest reds should go with it. Crowdin bridge green, 42 tests.

Thank you @Amoifr.

The rebase cleared both unit test reds as expected. Two things left here.

Psalm still reports the same two errors on getSupportedLanguageIds(), so the question stands: want me to push the cast, or would you rather fold it into your amend?

x86 / minimal-exts / lowest-php went red after the rebase, but it does not look related to this PR: #66040 and #66033 are red on that same job right now, each on a different AmpHttpClientTest timeout test (testTimeoutWithActiveConcurrentStream and testTimeoutOnInitialize, against testNotATimeout and testTimeoutIsNotAFatalError here). A different test each run on the same job, and the CI script already notes that HttpClient tests time out unless run separately.

I don't care about psalm if it doesn't spot bugs - so unless those reports are about bugs, it's fine not pursuing the goal of making SA tools happy for the sake of it.