Just to follow this up:

the problem seems to be with the order of 'preprocessing' steps in this case.
When I adwarp the whole functional bucket, and then use 3dbucket to extract the subbrik of interest, all three clustering options produce the same output.
However if the subbrik is chosen from native space first and then converted to TT using adwarp, the clustering comes out differently, and in the case of 3dclust not at all. I cannot think of a reason why this should be. It seems inconvenient to have to write the whole functional bucket to TT coordinates, and uses up too much diskspace.

Again, any ideas/explanations welcome.

Sarah