Interesting that it's hosted on GitHub. Everyone in this thread is wondering about the 2^-90 chance that a random commit will fail, but I wonder how these guys prevented their repo with test files from being rejected.
Filesystems like Venti could conceivably, on the one hand, check for collisions, but on the other, compare the stored block with the newly committed block. If there's a collision, but the blocks differ, just perturb the new block with a NOP, a nonce header of some kind...
Considering that the Bitcoin network burns through ~2^90 hashes in a year ... I'm not worried : )