Without adding significant amounts of noise to other WFS loops I have engaged the MC2_TRANS_PIT and YAW loops.
After several attempts to measure the system response and computing the output matrix, none of which gave any useful results, I gave up on that and decided to find three orthogonal actuation vectors which enable us to close the loops. So using the last good output matrix (below left side) as a template, I rounded it off to the nearest set of orthogonal vectors and arrived at the following matrix (right side):
I also decided that WFS1 and 2 need not drive MC2. This is just to decouple the loops and minimise cross-talk. This (albeit heuristic) matrix seems to work pretty well and the real matrix is probably quite close to it.
I show below the suppressed error signals after tweaking the gains a bit. The blue line is with no WFS, the green one with only WFS1 and 2 loops on, while the red is with all loops turned on. The WFS1Yaw and MC2_Trans_pit loops might benefit from a more careful study to determine a better output matrix.