Kiwamu found that MICH feedback signal suddenly becomes quiet about a second before the lock loss of PRMI.
Attached shows that this is due to the output of the M3 ISCINF filter (observed by the input of M3 lock filter, top right) creeping up and hitting the 5e5 count output limit.
Stefan remembers that this limit was necessary for lock acquisition to prevent one of OSEMs from going out of range and causing a bad alignment disturbance. Once the lock is acquired, then, the logical thing is to increase the limit or disable it.
Also, though the LSC feedback goes to the top stage, it seems like no optimization was done, M1 Lock filter only has 0.01 Hz pole. It should be possible to make a better blending to relieve M2 stage.