test_html_format rejects correct translations that reorder HTML elements (verb-final languages)

[openlibrary-bot]

Problem

test_html_format (openlibrary/i18n/test_po_files.py, and the copy in openlibrary-i18n tests/) compares a msgid's and msgstr's HTML trees tag by tag, in position. Languages whose word order differs from English legitimately reorder elements, and the test rejects those correct translations.

Found on #13704 (Azerbaijani): two correct translations reorder <b>/<em> and two <a> elements because Azerbaijani is verb-final. #13748 had to clear them to turn master green. In openlibrary-i18n, the pipeline's own gate keeps dropping the same class of string.

The better the translation, the more English we ship.

Decision needed

Relax the comparison to tag names, attributes and nesting as a multiset rather than by position? Or keep positional matching and accept the loss? Applies to both repos.

[openlibrary-bot]

Partly addressed on master by 18543f7 (@cdrini): test_html_format now accepts reordered elements when the msgstr starts with <!-- i18n-lint no-tree-order -->. Matching is order-insensitive but still multiset-exact, and there are doctests for both cases. The two Azerbaijani strings from #13704 carry the marker and are kept, not cleared.

What's still open:

• openlibrary-i18n's copy of tests/test_po_files.py doesn't know the marker, so a marked string would fail its gate there. The port (openlibrary-i18n#99) is handling it.
• Opt-in vs default. The marker is per string, so a translator or the AI pipeline has to know to add it. Nothing currently tells the translation prompt about it. The remaining decision: keep it opt-in (and document it for translators and the prompt), or make order-insensitive matching the default.

[MohitSAGAR11]

This came out of #13752, so some numbers from when I was diagnosing that.

Making trees_equal compare children as a multiset — sorting by
(tag, len(e), sorted attribute names) before zipping — took the az failures
from 3 to 1 locally. The remaining one was unrelated: a literal \x08 in the
spamwords.html msgid.

So the reordering case is cheap to fix properly. The question is the default.
The opt-in marker in 18543f7 works, but it asks a translator to know about
<!-- i18n-lint no-tree-order --> and remember it per string — and the people
most likely to need it are translating into languages furthest from English,
who are least likely to be reading our test source.

Flipping the default would make the common correct case pass silently while
still catching a dropped or invented element, which is what the test is
actually for. That sort key treats two siblings as interchangeable only when
they share a tag, child count and attribute names.

I would like to take this if the multiset default is the direction you want.

[kahinalrk]

Hello, can I take this issue to work on for a school project ? Thank you !