I saw the benchmark which seemed crazy (3us per transition and allocations). So looked quickly at the code.

Recommend putting the tracking of previous states in a separate optional debugging type so you don’t have to pay the cost in general. Oh and using time stamps as a key is kinda weird, but even weirder in Go where maps are non-deterministically enumerable.

Otherwise, seems like a good use of generics, given Golangs particular take on it.