The benefit of keeping the space there is that it provides a bit of space for the labels.

Expected:

Now:

I guess the bug it fixes is more relevant though and this is just a "co-incident" that it looks better with the old approach.

Can this PR get a bit more documentation - the all-black test doesn't help me understand what is going on, and the use of "Expected" and "now" above is ambiguous, and could use some test code to understand what was done and why. Thanks!

I think I'll have to ask @anntzer for a proper explanation as I just implemented his patch. For now, I'll have to refer to the issue it closes.

In current main, there is a thin white line to the left and below the figure:

The "expected" image is the current test image. The "now" image is the result after this change.

My interpretation is that we remove some non-visible ticks from the size calculation which gets rid of the white stripes but also breaks the nice looking alignment on the x-label for the color bar.

Edit: reading the code I've updated the description based on what I believe happens (at least on a higher level, there is a bit of Python magic that I do not really follow).

Your interpretation is correct. The first part of the patch skips any handling of the axis if it is not visible. The second part filters the visible ticks, and includes only the first visible major tick (if any) and the first visible minor tick (if any) into the bbox calculation.