I haven't checked locally, but at least CI doesn't seem to be agreeing with you... (you can run the tests locally with run-mpl-test-suite.py; if the fix is correct you should be able to remove the excluded tests from the hardcoded list in that file.

I realized that whatever I have done in this PR is not correct for sure. I am going to spend some time on this over the weekend... looks like I will have to dig into mpl code to figure out all the magic you described

The main problem I am aware of is with using figimage() (so "per-pixel data") with pdf output (where the notion of "pixel" is not really well defined, but conventionally thought of as 72 dpi); I think there's somewhere where I may forget to switch the canvas from 100 dpi (or whatever the "nominal" resolution is) to the pdf output's 72 dpi, or where mpl does it and I do it too so it happens twice (IIRC (from very long time ago...) mpl's image.py code basically does something slightly weird to match the just as undocumented behavior of backend_pdf.py, and I gave up reverse-engineering that).