I tested 2.1.0 against LexicRo, which uses verbecc for Romanian.

Conjugation output: no regressions. I dumped CompleteConjugator(lang="ro").conjugate(v).to_json() for all 6,869 Romanian verbs on 2.0.3 and on this branch. Every one is byte-identical, even with the move to dataclass-wizard 1.0 and scikit-learn 1.9. The only difference is pred_score on predicted verbs (e.g. blogui 0.9978 → 0.9959), because the model is retrained under the newer scikit-learn. The chosen template and forms are the same. The #49 fixes survive the move, and #50 and #53 are unchanged, as expected.

Two things need fixing before release, though.

1. The built wheel contains no data (blocker)

pip wheel . on this branch gives a 181-file wheel with none of verbecc/config/.yaml or verbecc/data/xml/**/.xml. A normal install fails on first use:

ModuleNotFoundError: No module named 'verbecc.config'

There are two causes:

[tool.setuptools.package-data] is keyed trained_models, which isn't a package, so the globs match nothing.
config/ and data/ have no init.py, so packages.find skips them.

CI doesn't catch this because it uses pip install -e ".[dev]", which reads the source tree directly. This fixes it:

[tool.setuptools.package-data]
verbecc = [
"py.typed",
"config/.yaml",
"data/xml/conjugations/.xml",
"data/xml/verbs/.xml",
"data/models/",
]

A CI step that builds the wheel, installs it into a clean venv and conjugates one verb per language would prevent this happening again.

2. from verbecc import CompleteConjugator no longer works

src/verbecc/init.py is now empty, so the entry point shown in the README raises ImportError, and every existing consumer breaks. The changelog flags that some imports may break, but this is the main public entry point, not an internal path. It's worth keeping, and it can be kept without breaking R006 by re-exporting one level at a time:

src/verbecc/init.py

from .core import CompleteConjugator as CompleteConjugator

src/verbecc/core/init.py

from .conjugator import CompleteConjugator as CompleteConjugator

src/verbecc/core/conjugator/init.py

from .complete_conjugator import CompleteConjugator as CompleteConjugator

This loads nothing new, since CompleteConjugator pulls in the whole conjugator anyway, and there are no circular-import problems. If you'd rather make the break on purpose, please document the new path (verbecc.core.conjugator.complete_conjugator) in the README and in the changelog's breaking-changes section.

With both fixes applied, the wheel installs cleanly, all six languages conjugate, and the test suite passes (556 passed, 2 skipped).

Thanks, Brett, and no apology needed.

Re-tested at 412f3f8:

Both blockers fixed. The wheel now ships the data and config, from verbecc import CompleteConjugator works again, and the wheel smoke test in CI should keep it that way.
Romanian output unchanged. All 6,872 Romanian verbs are byte-identical to 2.0.3, from a clean-venv install of the built wheel, apart from pred_score on predicted verbs.
Suite and lint clean. uv sync --locked, ruff and pytest all pass (558 passed, 2 skipped).

Copilot’s xml_utils.py:31 finding reproduces: str(etree.tostring(...)) returns the text of the bytes literal, not decoded XML. In the library it only reaches error messages. scripts/xml_modifier.py does need bytes, though, and currently fails with TypeError on write. Having elem_tobytes call etree.tostring directly, and xml_element_to_string return .decode(encoding), should fix both. It’s not a blocker for me, but it’s worth fixing before release.

Approving from my side.