Came in via 7898bbb / #2147

The test failures look real (likely in the anti-alias of the line along the edge).

I spent a while hunting this down until I realized that, in fact, the before- and after- images should look different. For the moment I can't figure out how the old test conspired to make clip and mask so similar.

One possibility is the way MPL chooses the axis bounds. Shown below are four plots, the Cartesian product of (clip, mask) and data which starts at x= -2 and x=0.

compare the above output from mpl 1.5.1 to this from mpl 2.0.0:

is is evident that mpl1 has chosen x-axis limits that mostly obscure the difference between clip and mask.

It seems that whichever behavior we want, we will need a test that more explicitly checks the distinction.

Most old tests still use the classic style, where the plots use "round" numbers as limits, so that's probably why they still passed. Updating the test images should make this good.

This ended up being superseeded by #8836, thanks for the PR though!