A lot of the intermittent failures ("oranges") that are happening on the test machines, are race conditions, like low memory conditions/high disk usage/slow web server/etc.
One thing that could be improved is to clear the cache after each mochitest. But I also think the tests themselves shouldn't rely on bfache to work (unless of course, they are testing bfcache). So I filed bug 1057499 and pushed a tryserver run there to see how things go with bfcache turned off. Could be interesting.