I assume what the AGA is doing, is that it has some numerical rating coming out of a model, and then it scales in a way that it can assign integer labels that we interpret as a kyu dan system.
All in all though, ranks are just fancy labels of ratings, unless they connect to some aspect of the game. You could call someone 2000 or 1d and it makes no difference, the same way you could call them brown belt.
The extra freedom you tend to have in the kyu-dan system is that when you want to assign an interval of ratings to a specific bucket like 1dan or 2kyu etc, you now have a parameter which is the size of the interval or the bucket.
If you make that interval a constant width, either in the original ratings or after some linear scaling or something like that, then I agree that maybe 20vs23kyu and 2d vs 5d might not have the same win distribution.
However you allow the intervals to change width as a function of ranks, then you can account for this difference. You might have small intervals for tpk or ddk players and wider rank intervals at high dan, to accomodate for the difference in skill needed to win with extra stones for example.
I noted this about OGS previously here
Again I don’t really know what the AGA does, so maybe you’re generally correct about your assertions about its system.