Lie about immediate effect, wait 1.5sec on press A. If it was <1sec press, proceed. Otherwise, wait for another 1.5sec and then work on B modality.

It depends on the lie, but I think it probably fits human-centric timescale.

The alternative is to work on the partially ordered set of events {A, B, undo-A} and convert to {A, undo-A}, {B} and then transform to {no-op, B}

Hysteresis/Queue is your friend.