Perhaps do a test like given in #16461 but settings x/ylim so that in the correct case the entire patch is out of bounds, but would be drawn at (0, 0) otherwise, and then check_figures_equal() against an empty plot?

That worked well - the test I've added fails on current master but passes here.

If we need to change baseline images (especially non-hexbin-related ones) I'd really rather just go with #13696 -- this bug has apparently been present for a long time, I'm not convinced it's RC to have it fixed in 3.2. Or if it helps I can make VectorTransform public in #13696; I don't like it but it's not the end of the world.

Agree with not needing to put this in 3.2. Won't the baseline images have to change anyway at some point though?

I don't think so. I think (not 100% sure?) here you disabled single_path_optimization when offsets are nonzero, but they really only need to be disabled when offsets are nonzero and offset_position="data". When offsets are nonzero and offset_position="screen" (which will always be the case after #13696), the old baseline images are still fine.

Ah yes, true - I can check for offset_position = 'data' here and not worry about the baseline images then.

Surely going through either of the two code paths should give identical results, but it doesn't look like it does because the images have changed. I guess the question is for offset_position = 'screen', which code path is doing the right thing?

Basically single_path_optimization is slightly "wrong", because it "rounds" the markers to the closest pixel (it's fundamentally the original reason I wrote mplcairo...), see e.g. #7233.
But there are cases, e.g. ticks, where the rounding is actually useful because if you try to draw short ticks antialiased "between" two rows of pixel, they look particularly bad.

So this is replaced by #13696?

I think so? feel free to reopen if not.