Still, some numerical instabilities in PGVA diagonal weights update when too many components are used. I think it happens when the smallest eigenvalue is smaller than the residual noise.
LB still drops in PGRA when updating the residual field, despite some fixes. I'll keep investigating.