A fresh test run on French exposed a real bug in _apply_translation:
when the model returns a JSON list for a plural entry (e.g.
["form0", "form1"], which is a valid representation since plural forms
are ordered), the previous code took the else branch and broadcast
str(list) — Python list-repr like ['form0', 'form1'] — to every plural
form. Both msgstr[0] and msgstr[1] ended up containing the same
literal Python list-repr string, breaking gettext lookups for that
entry. Spanish dodged it by chance (the model returned dicts that
time); the failure mode is reproducible on French.
Changes:
- Extract _apply_plural_translation helper. Handles dict, list,
scalar, and non-JSON-string responses. List path distributes forms
by index and repeats the last form if the model returned fewer
forms than the language requires (better than leaving slots blank,
which falls back to displaying the raw English msgid).
- The split also drops _apply_translation's cyclomatic complexity
back below the C901 threshold.
- Adds 4 regression tests covering: list response, list response
round-tripped through parse_response, list shorter than required
forms (last-form-repeats), and empty list (falls back to raw-string
broadcast).
Verified end-to-end on French: the previously-broken plural entry
"Added 1 new column to the virtual dataset" / "Added %s new columns
to the virtual dataset" now writes msgstr[0] and msgstr[1] correctly
on a fresh run.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Addresses sadpandajoe's review on #39448:
1. Adds tests/unit_tests/scripts/translations/backfill_po_test.py with
19 cases covering parse_response (singular/plural/markdown-fence
stripping/non-ASCII/non-numeric keys/list-and-scalar rejection/JSON
errors) and _apply_translation (singular path, plural-dict path,
plural-scalar fallback, plural invalid-JSON fallback, fuzzy flag,
attribution append/dedup, end-to-end round-trip from parse_response
into _apply_translation). The script is loaded via importlib since
it lives outside the package tree.
2. translate_batch now pipes the prompt over stdin instead of passing
it as argv. With --batch-size 50 and many reference languages a
single batch can grow into the tens of KB and approach ARG_MAX on
some platforms; stdin removes that ceiling.
3. _process_batches now saves the catalog after each batch that wrote
at least one translation (when not in --dry-run). For sparse
languages with thousands of missing strings, a crash mid-run now
only loses the in-flight batch rather than every batch translated
so far. The full save at end of backfill() is removed since the
per-batch save covers it.
4. Module docstring referenced --fuzzy/--no-fuzzy but argparse only
registers --no-fuzzy; doc updated to match the actual flag.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
