[jbri]

No, that's not it.

If you restart the simulation a few times, you should eventually get something where everything forms into one clump with low initial angular momentum. If you keep watching from that point, you'll see the clump's rotation accelerate entirely on its own.

I'm not entirely sure on the cause, but it does seem to be a simulation flaw.

[drakeandrews]

You're right. At a guess, as collisions appear to be perfectly elastic particles inside a clump are trying to move and instead of their shifting being dampened, it's causing the whole structure to begin to rotate which causes more in-clump collisions causing... until it spins apart. Sound reasonable?