I think a test that writes the empty figure to a BtyesIO buffer and either checks that font is not there in the postscript or just that it is small enough would be enough.

Yes, that is indeed one way to test it. I was actually more worried about breaking something than still producing large files, but I'll add a file size test.

There is now a test. However, the figure size selection logic before distilling is not really obvious to me. It may be that it is just needed for distilling as the resulting "paper size" is the same here for both ps and eps (the figure size), but there are for sure cases where this doesn't happen, like when the figure is larger than letter for ps. So I'm a bit doubtful if is may break something else...

Not distilling leads to that ps-images are no longer in "paper size", but in figure size. I cannot really see where this is a problem, but since this is breaking current behavior, I put this in draft stage and set it up for discussion on the dev call.

There is a quite simple fix for this, determining paper size, before writing the header, but we should probably discuss what the "correct" behavior is for different cases.

FWIW I think we should eventually kill the papersize kwarg and (if we want to keep an equivalent feature) let users specify e.g. figsize=mpl.get_papersize("a4") (or even figsize="a4" if we really want that conciseness) in the Figure constructor, to make things more regular across backends (and avoid e.g. mplcairo having to reproduce that logic).

As it seems like @QuLogic is doing an overhaul of the PS backend, I tag you in.

IIRC the discussion at the dev-call was about if someone really needs ps output anymore or if we should go all eps. This is combination with @anntzer's good remark about the use of the papersize argument above should probably be the subject of a new dev-call to possibly make a decision on the subject.

Possibly related: #22416 as it involves the eps/ps distill chain as well.

I think in principle, this makes sense, but I guess I need to do some tests to see what the exact results are.

Yes, I do not recall exactly why I put this into draft. I think it was because I did not really get the full sequence of distill-runs. But I guess that, if possible, it would make sense to avoid embedding if not used.

I also know that I wanted to add a test to make sure that one actually gets EPS or PS. Maybe I mix up these topics... Or the issue here was that unless we distilled the "EPS" would end up to be a PS. Or something...