The problem has been solved. It had to do with differences in the implementation of thresholding in the clustering code and in the main AFNI GUI. The fix is explained below, in an excerpt from README.environment (as I edited it yesterday). However, you'll have to wait until we re-compile the binaries. I hope to remember to do that tonight, but my aging brain doesn't remember things as well as it once did :(
-------------------------------
Variable: AFNI_OLD_SHORT_THRESH
-------------------------------
When thresholding a dataset with a sub-brick that is stored as shorts (16 bit
integers), the AFNI GUI uses floats, but the 3dmerge and Clusterize functions
use shorts. The difference is that the user-supplied threshold in the latter
case is rounded to the nearest short. Thus, a threshold of 2.2 would become
2, and then a value of 2 would pass the 'greater than or equal to threshold'
test -- which is probably not what the user meant. Again, this would happen
in 3dmerge and Clusterize, but NOT in the AFNI GUI without Clusterize.
This inconsistency has been fixed, and both sets of places now threshold
using floats. However, IF you want to stick with the old method for some
un-imaginable reason, you need to set this variable to YES.