I've been using r31446 all day, and decided to give the acid3 test another whirl this afternoon (to stare at that 100/100 glory), but I was shocked to see test 79 fail the first time around:
Test 79 failed: character position 6, which is between a normal character, which has some kerning after it, and the first character of a two-character glyph, which has some kerning before it, is 21600 but should be 21800
Immediate reload and it went back to 100/100. Took screen shots of both states (Before and after reload).
I'm pretty sure we have a dupe of this. I think Acid3 actually has timing issues. It doesn't force the font load to be synchronous.
I think we fixed that a while back. Can't reproduce.