As of #29034, _weighted_percentile handles NaNs by ignoring them when calculating percentile.
np.median and np.percentile on the other hand, will return NaN if a NaN is present in the input (np.nanmedian and np.nanpercentile will ignore nans).

There are many cases in the codebase where, if sample_weight is None, a np function is used (NaN returned), if sample_weight is given, _weighted_percentile used and NaNs ignored.

Summary of affected cases:

• DummyRegressor.fit
• AbsoluteError/PinballLoss/HuberLoss - fit_intercept_only method
• median_absolute_error
• d2_pinball_score
• SplineTransformer._get_base_knot_positions - I think this was the original reason for MNT _weighted_percentile supports np.nan values #29034

Maybe we could assess on a case by case basis whether it makes sense to return NaN if present in the input? @ogrisel suggested that we may want to raise a warning in some cases as well.

cc @StefanieSenger