I think I have the solution. I segmented both the template (yes
to "spatially normalized") and the source (no to "spatially
normalized"). I then used the GM seg-image from the template as the new
target and the GM seg-image from the source image to do spatnorm. Re-
applying the *_sn.mat" to the original source T1 (and its co-registered
PET image) gave much much better results.
If anyone has another solution, that's fine, but I think I'm OK.