This exposed some problems in our clipping logic, which was never
properly rewritten for the stacking context reform. The clipping code
worked in terms of a stack of clips, but the new stacking context code
has no concept of a stack of clip regions. Fixing that in turn exposed
some flaky/incorrect tests:
* `borders` had an incorrect reference image, as far as I can tell.
* `negative_margins` had some stray pixels, fixed by changing the text.