The proposed changes seem equivalent (potentially slower?)

From my understanding: If the training data, X_binned_train, has no missing values, we would not need to do the right to left sweep when splitting.

Correct. When there are no missing values, we only do one scan (left to right).

Secondly, this seems leaks data from the validation set, X_binned_val, into training.

How? has_missing_values is computed with X_train, not X.

How? has_missing_values is computed with X_train, not X.

Ah yes, I was mistaken, this PR is not strictly needed.

The proposed changes seem equivalent (potentially slower?)

Although this change is faster:

from sklearn.ensemble._hist_gradient_boosting.binning import _BinMapper
import numpy as np

rng = np.random.RandomState(42)
X = rng.rand(100_000, 200)
mask = rng.binomial(1, 0.1, size=X.shape).astype(np.bool)
X[mask] = np.nan

mapper = _BinMapper()
X_trans = mapper.fit_transform(X)

%timeit (X_trans == mapper.missing_values_bin_idx_).any(axis=0).astype(np.uint8)
# 2.31 ms ± 69.6 µs per loop

%timeit np.isnan(X).any(axis=0).astype(np.uint8)
# 14.4 ms ± 394 µs per loop

Secretly, I am using this change for categorical support and this PR may be an improvement (at least in terms of speed).

I guess we can merge without a +2?

Looks okay to me.

Reverted this from master in c9ca1d7 since it broke master

It was broken because of #13022 which does not work for new versions of sphinx.

Will be fixed with #17724

I don't think so: https://dev.azure.com/scikit-learn/scikit-learn/_build/results?buildId=18715&view=logs&j=13650d7c-49e8-54f3-0598-c3480d1c1e4f&t=97f2162d-a0c4-54dc-9fce-2fceef7172e3&l=40756

Ah okay I see.