Propagate NaN from DataFrame.skew and DataFrame.kurtosis - #12584

skew() and kurtosis() end with

result = m3 / m2**1.5
if result.ndim == 1:
    result = result.fillna(0.0)

result.ndim == 1 is true exactly when the input was a DataFrame, because the reduction of a DataFrame is a per-column Series. A Series input reduces to a 0-d scalar and skips the fillna entirely. So the two paths disagree: for a column holding a single NaN, Series.skew() returns NaN and DataFrame.skew() returns 0.0, and kurtosis returns NaN and -3.0 respectively, from the same data in the same library.

Both docstrings say the method does not filter NaN, and nan_policy defaults to "propagate" and is the only accepted value - "omit" raises NotImplementedError. 0.0 and -3.0 are therefore values the documented API says it will not produce. They are also both inside the plausible range for real data, meaning "symmetric" and "strongly platykurtic", so nothing downstream looks wrong. On a 200k-row frame, one missing value in one column is enough to turn a skew of 6.83 into 0.0 and a kurtosis of 155.8 into -3.0.

dask.array.stats.skew, which the docstring says the implementation follows, and scipy.stats.skew under nan_policy="propagate" both return NaN here.

The existing tests only exercise the Series path, which is the path the fillna skips, so nothing covered this.

Note on constant columns, since a reviewer will ask: for a column of identical values pandas returns 0.0, scipy returns NaN, and dask's Series path returns NaN. This change makes the DataFrame path return NaN, so it agrees with the Series path and with scipy but not with pandas. The NaN-bearing column is the unambiguous case; the constant column is a consistency judgement and I am happy to special-case it if maintainers prefer pandas parity there.

• Tests added / passed - new test_skew_kurtosis_propagate_nan_on_dataframe passes, and the full dask/dataframe/tests/test_arithmetics_reduction.py suite is green (173 passed, 3 skipped, 4 xfailed)
• Passes pixi run lint - ruff, black and mypy all clean on both changed files