Yes, i was vaguely aware of what was going on in the mysterious auto-tlrc script..
For the reason you mentioned, I would prefer to use the "first method", getting all the transformations into one step. My concern is that when I combine all the individual steps that I did, the result is not the same.
Perhaps these pictures will explain..
photo one shows my "template" registered to a talairached structural.
steps performed were:
functional-to-functional registration (3dvolreg)
functional-to-structural registration (3dAllineate)
apply talairach transform matrix (auto-tlrc -apar)
the resulting alignment is rather satisfying
picture two shows what happens when i combine the 3 transformations acquired into 1 and apply it to my original data. If you look closely, you may or may not be able to see that the functional image is shifted to the left by about a voxel or so (4 mm)