uk.json declares direct_address_patterns (plural, list of 10 regexes) under entity. The loader at mempalace/i18n/__init__.py:209-210 only reads the singular key direct_address_pattern as a single |-alternation string; the plural key is never consulted. Every working locale (en, ru, it, pt-br, hi, id) uses the singular form. Effect after merge: all 10 direct-address patterns in this PR are silently dropped, and Ukrainian users get no direct-address entity detection until the schema matches.

Fix is collapsing the 10 alternatives into one regex and renaming the key:

"direct_address_pattern": "\\bпривіт\\s+{name}\\b|\\bдякую\\s+{name}\\b|\\bдобридень\\s+{name}\\b|\\bдобрий\\s+день\\s+{name}\\b|\\bшановний\\s+{name}\\b|\\bшановна\\s+{name}\\b|\\bдорогий\\s+{name}\\b|\\bдорога\\s+{name}\\b|\\bhey\\s+{name}\\b|\\bhi\\s+{name}\\b"

Minor, separately: two person_verb_patterns have a literal hyphen in their alternation group:

"\\b{name}\\s+засмія(-|в|л)(ся|ась|ись)?\\b"
"\\b{name}\\s+посміхну(-|в|л)(ся|ась|ись)?\\b"

(-|в|л) matches a literal - or в or л. Natural past-tense forms are засміявся / посміхнувся without a hyphen, so the intended group is (в|л).

Otherwise topic_pattern, quote_pattern («»), candidate_pattern, and the stop_words / stopwords split align with ru.json. Once the direct-address key is corrected the file should load cleanly.

@mvalentsev sorry, missed that.
thank you for checking.

Thanks for the Ukrainian locale — content is structurally solid, and the hybrid Cyrillic+Latin candidate_pattern is a nice touch for UK/English mixed tech prose. Two small things before I can merge:

1. Drop multi-word stopwords. "будь ласка" in the stopwords list can never fire — the BM25 tokenizer splits on whitespace (\w{2,} tokens), so space-containing entries don't match any single token. Either split it into separate single-word entries or remove it. Same for any other multi-word phrases in the list.

2. Add a uk sample to tests/test_i18n.py::test_dialect_compress_samples — every other shipped locale has one, and it's a useful smoke test that Dialect round-trips your locale correctly. Something like:

   "uk": "Ми вирішили перейти з SQLite на PostgreSQL для кращого паралельного запису. Бен учора схвалив PR.",

I'll kick off CI once you push the fixes.