Hi-
How about adding this option, to have lpa used as the cost function for the final rounds?
-cost_nl_final lpa
Note that the init_qc_01_nl0 is not the final nonlinear (NL) result, but the start of the main nonlinear part. That means that the first image there is showing that there will be a bit of ways to go. It is possible that the lpa cost, above, will help resolve the alignment at the later stages.
I woudl be curious to know how the anatS.*.nii volume looks? Sometimes the initial 3dSkullStrip result can be poor, causing later issues.
--pt