I hope that this is largely an academic discussion and no one is actually suggesting to implement backlash both ways. Backlash removal is an ancient concept and at the core is the choice to make a measurement by approaching from the same direction every time so that the readings are consistent. Often there is a preferred mechanical direction to approach from that has less compliance, slop, and variability - and in the case of a focuser it is the direction that is pushing up against gravity. So you estimate how much slop there is - and as long as you make a move in the good direction that is greater than the slop amount, you know the reading will be good and repeatable - no matter where you started from.
Imagine having this bi-directional system set up where you have some presumed value for the slop and you actually use that value when moving from each direction. Then imagine you make 100 moves of one step in alternating directions. If you have a slight error in that estimate of slop, that error will accumulate with every small move - and you will have no idea where you are.
Backlash removal works very well from one direction - especially if it is set up in the optimal direction that has less slop. And it only kicks in when you need to go backwards - which during autofocus only happens at the start of the run, and at the end of the run when it returns.
The only thing needed in setting it up is getting the direction of the curve right - which is usually pushing up against gravity as the curve is taken right to left - and the backlash amount, which can be determined by feel or by watching the system focus after a change of direction. Too small a backlash is bad, but using a large value is probably fine as long as it isn't too big - and you can always tell if things are working if the focus curve is repeatable.