I see where things are confused. In this case, I guess it depends on whether you store the entire sequence of states of the application or if you're storing the entire sequence of events emitted.
I was thinking of the former, while you were thinking of the latter. If the entire state of the application is contained in one place, you don't need to reply events from the beginning. You just need the application state at the point.
Purity has little to do with implementing these kinds of debuggers, effects need to be explicit, but this can be dynamic explicitness rather this static kind.
The question can be interpreted in multiple ways, but this is not a feature novel of Elm (see "time steering" in the 90s...and debuggers like xstep).