[alejandroschuler]

Found while assessing #168 (PR #433).

test_dists_runs_on_examples_logscore in tests/test_distns.py had this marker:

@pytest.mark.xfail(
    condition=is_t_distribution, reason="Known to fail with T distribution"
)

is_t_distribution is a function, and pytest reads a non-string condition as a truth value. A function object is always true. So the test was marked xfail for every distribution and every learner, not for T alone. A real failure in Normal, Gamma, Weibull or any other case would have shown as xfailed and passed CI. With --slow, the 22 passing cases showed up as xpassed, which is the visible sign of the problem.

Change

• The T case in the dist parametrize list is now pytest.param(T, marks=pytest.mark.xfail(...)). The marker on the test function and the is_t_distribution helper are gone.
• The reason names the failure and cites Issue with T distribution #275. NGBRegressor(Dist=T) with default settings raises ValueError: Input y contains NaN from fit, because the natural gradient for log df runs away as df grows. np.exp(params[2]) overflows, and the NaN reaches the base learner.
• The marker uses strict=True, because T fails for both learners. If T starts to pass, the run fails and the marker has to be removed. I checked that the failure is the same on Python 3.9, 3.12 and 3.14 (the ends and the middle of the CI matrix), so a strict marker should not flip on one CI job.
• The marker uses raises=ValueError. A different exception from the T case now counts as a real failure.

I checked the rest of tests/ for the same pattern. This was the only xfail. The skipif conditions in test_helpers.py and test_pickling.py call their helper (not _sklearn_has_missing_go_to_left()), so they are real booleans.

Tests

All runs use Python 3.12 and scikit-learn 1.6 with --slow.

tests/test_distns.py::test_dists_runs_on_examples_logscore, 24 cases (12 distributions, 2 learners):

Case	Before	After
T, both learners	xfailed	xfailed (strict, raises=ValueError)
the 11 other distributions, both learners (22 cases)	xpassed	passed

No case fails after the change. No distribution other than T was hiding a failure.

Full suite:

	Before	After
passed	168	190
xfailed	2	2
xpassed	22	0
skipped	1	1
failed	0	0

The before counts come from two runs on an unmodified export of origin/master: tests/test_distns.py alone (28 passed, 2 xfailed, 22 xpassed) and the rest of the suite (140 passed, 1 skipped). The after count is one run of the whole suite.

Lint on tests/test_distns.py is clean: black --check (24.10.0), isort --check-only, and pylint with the options from .pre-commit-config.yaml (10.00/10). flake8 in that config only covers ngboost/.

🤖 Generated with Claude Code