Could easily be back ported if not a bother.

Should a non-visible axes take part in the layout? If the answer to that is yes, this is a good fix; but one might also consider that a non-visible axes may just be treated as "not there", or maybe as the most tiny box possible at the center of its original position?

Forget the above. Yes, this is the correct fix. If you want to not have the axes take part in the layout you should use.remove() instead.

An empty gridspec box gets a placeholder like this anyhow, so this is the smallest box that is possible for this spot if gridspec orsubplots has been used for the layout.

I do this because if someone has a gap in their layout presumably they want it to be there. And it is needed to keep the axes in rows above or below a blank (for instance) aligned.

This is somewhat poorly documented at:https://matplotlib.org/3.1.1/tutorials/intermediate/constrainedlayout_guide.html#empty-gridspec-slots
and its checked in one of the tests... (the horribly named:test_constrained_layout8)

Yes, I realized that 5 minutes after I posted (see edit above), sorry for the confusion.

Too fast! Though I'll point out that even remove will still put a blank gridspec in there (actually an axes with all the elements made invisible!)

flake8:

./lib/matplotlib/_constrained_layout.py:279:1: W293 blank line contains whitespace

Hmmm, I kept trying to rebase that change in, but I suspect I added the space in the "TST" commit, so it kept overwriting it. Ooops....