Unrelated to this issue but I've had a race condition with Automake which while run oin 2-4 threads occured exactly every 2nd run. With -j48 it was obvious it's a race condition. No idea how cache invalidation works in the automake stack, but that must have caused it to fail exactly 50% of the time.