First off, I am not sure the lag compensate mechanic is applicable to this situation.
It seems to me the required mechanic here relies on knowing where a paddle/ball might be, in stead of having to know where it was.
But, seeing as there is a level of predictive-ness involved in the plugin, it might apply.
I have not made a mechanic like this yet, but I would start by trying:
if the ball passes a certain Y , on either end where it crosses a line where it 'could' interact with a peer's paddle, I would perform a quick lag compensated X,Y on the peer's paddle, check for collision with the ball, then return paddle to original position.
Remember, only the host can perform the lag compensate re-positioning. So if the host deemed it a collision, it will perform the actual collision.
Right after, you replace the paddle back to its original position, but it will have performed the bounce if that is already in the host mechanics. (and if it really works )
I always envision the lag compensate test as a method to determine a connected users position (either host or peer), a moment in the past based on the lag time from a peer(!not host).