While I definitely prefer the approach where you grab one or more steps first, and then apply a function (rotate a knob, press a button), the main problem right now I have with that approach is that the steps involved are toggled off the moment you grab them, and ‘fixed’ again the moment you apply the function. So if you wait a little too long between grabbing and applying the function, the composition is changed (no notes).
When grabbing a step, the step should not be toggled off on the key-down event, but on the key-up event instead, but only if during key-down and key-up no other key or rotary event was sensed. Something like that.