When I read "wasn't entirely correct", and "probability of a crash was about one in a million", I wondered how you'd work that value out.
Predictable statistics screams RNG, to me. Especially if the code author knew it was an imperfect solution, and didn't know of one that wouldn't cause this crash. (I've done stuff like this before.)
Also, for the client to crash the server, it'd need to give it some form of bad/confusing data, but for some reason this doesn't happen with fewer users.
I think that was my logic at the time of writing that comment, though I've thought of another few guesses while typing this out/reading your follow-up.