It seems to make assumptions about text positioning that are not holding with subpixel positioning. I'm not 100% sure how that leads to exactly the artifacts that are seen here, but I am just disabling the test until that is fully understood.