Somehow, I generically don't like the idea of lead filters for the WFS loops. We don't really need so much bandwidth. I think you should include with the servo measurements, a servo model ( on the same plot ) that matches the loop shape.
For example, this means including the 28 Hz ELP in the MC1/3 hardware and MC2 ASCPIT/YAW digital filter banks. BY comparing the model v. measurement we can determine if the cross-coupling due to imperfect output matrix is very serious or not.
In the measurements, the loop with the most low frequency gain looks the most promising.