Not obvious why this kills all Linux tests, but no Windows or Mac tests.

This looks extremely fragile. After a rebase I get even more errors... Especially, it seems like the conversion from PDF to PNG is affected for some unknown reason.

An example:

Screenshot of expected PDF

Screenshot of resulting PDF

These look quite the same IMHO.

PNG of expected PDF

PNG of resulting PDF

Expected PNG for reference

I cannot run the whole test suite locally in one go as I get a Fatal Python error: Illegal instruction every now and then. Most tests that fails here pass locally though (some minor, single pixel type of errors).

However, I noted this error from the test log:

! Package geometry Error: \paperheight (0.72229pt) too short.

See the geometry package documentation for explanation.
Type  H <return>  for immediate help.
 ...                                              
                                                  
l.1006 \Gm@process
                  
No pages of output.
Transcript written on figure.log.

and the following error:

So it may indeed be related to what was discussed during the dev call: something messes up the ghostscript conversion process that propagates among the tests.

Fatal Python error: Illegal instruction

Those tend to come from binaries compiled with CPU instruction sets that your CPU does not have...and I have mostly seen them with mkl or other "go fast" numpy extensionss....

Hi @oscargus any plans to finish this up? Particularly so that #29392 can get rebased on it?

I'd be happy to finish it. Just that I do not really know what remains to be done and/or what causes the weird inconsistent test images.

But I'll try to rebase it and see what the current status is and if anything has changed.