One way to avoid this problem is to use IMAGE BYSEG for the micelle, so that it remains intact throughout the simulation.
Use the LONGline command prior to writing the data files from CORREL to avoid line wrapping to 80 chars.
I'd probably look at time series of the P atom coordinates to see if I could find one that hasn't been image centered at all, and use that as the target for the RECENTER.
Have you tried w/o ORIENT NOROT yet?
I have resolved z drift in some bilayer simulations by using the COM from the unfolded trajectory; however, this requires starting from a point before any lipids have been separated by image centering.