Having done some trial and error using compression on recorded drums I really didn't know how it all worked. But this explains it quite well:
Also it is now clear why they ask for consistent levels on each hit when recording: the compression threshold has to be in exactly the right place for it to work. Not a musical but engineering requirement.
But maybe now in the digital age there should be some kind of adaptive compression that puts the threshold in the right place regardless how loud the stroke is?