On the timescale of stimulus-based action–effect learning