[biodranik]

determine_language_given_path() falls back to scanning every segment of the path it is given for ^[a-z]{2,3}(-[A-Za-z]{2,4})?$, and generate_all_localization_files() handed it the joined absolute path of each candidate folder. Any short lowercase ancestor of the checkout matches:

/private/tmp/omim/iphone/Chart/Chart/ChartData   -> "tmp"   misfire
/home/ab/omim/iphone/Chart/Chart/Views           -> "ab"    misfire
/opt/build/omim/data/sound-strings/en_GB.json    -> "opt"   misfire
/Users/alex/Developer/omim/…                     -> None    ok

So every non-language subfolder of an output directory was treated as a language. With --include all that wrote full English files into source folders; with --include translated the output was empty and silently skipped, which only hid the same misclassification. Worst of all, generated output depended on where the repository happened to be cloned.

Both generate_all_localization_files() and find_translation_files() iterate the direct children of a single directory, so the language can only come from the child's own name — ancestors are meaningless there. Pass the name instead. _prepare_read_write() is left alone: it receives an arbitrary file path such as …/en.lproj/Localizable.strings, where the segment walk is the point.

That alone still left folders named raw and xml matching. Formatters with a fixed convention now return None rather than falling back to the generic regex: an Apple localization folder is always <lang>.lproj, an Android one always values[-lang].

The same flaw is in lib/twine/runner.rb; changed there too for parity.

Testing

Running organicmaps' tools/unix/generate_localizations.sh from a checkout under /private/tmp, against a clean tree:

	before	after
working tree after generation	1 file overwritten, 2 bogus files created	clean
files generated	675	672
"would not contain any translations" skips	68	6 (all genuine)

The 672 legitimate files are byte-identical to what is checked in, and a second run is idempotent.

pytest: 134 passed, including two new regression tests. One drives the runner through a TemporaryDirectory, which reproduces the original bug naturally — macOS temp dirs live under /var/folders.

[biodranik]

Rebased onto #31, which fixes the pre-existing Ruff findings described above, so this now targets ab/ruff-fixes instead of organicmaps. Merge #31 first and GitHub will retarget this back to organicmaps automatically.

The only conflict was the import block in test_integration.py, where #31's isort pass and this branch's new Runner import met. pytest passes 134 (132 + the two regression tests), ruff check is clean, and organicmaps' generate_localizations.sh still emits 672 files with 6 genuine skips and a clean tree.

[strump]

LGTM