As part of this bug, we also need to revert the patch landed in Bug 1295466 that uses MozReftestInvalidate event to stop the test-failure from showing up when reducing paint delay.
Update: we rolled back the change described in Comment 1 and we instead added an extremely-permissive "fuzzy" annotation, here: https://hg.mozilla.org/integration/autoland/rev/c46aadcf745c#l2.12 So now, when this bug is fixed, we'll want to make the following changes on top of the patches that landed in bug 1295466 comment 46: (1) Remove that ^^ fuzzy annotation when this bug is fixed (2) Backout https://hg.mozilla.org/integration/autoland/rev/faa9981549a6 (i.e. remove the duplicate files that we created there, since they only exist to allow us to run a version of this test with a Mozilla-specific "MozReftestInvalidate" hackaround)
I think the test failures in bug 1295466 are caused by this same underlying issue, too. (At least, the failure snapshots look similar, with the right chunk of a scaled image missing. E.g. in https://treeherder.mozilla.org/logviewer.html#?repo=mozilla-inbound&job_id=75524537&lineNumber=30184 )