Dear Andrew,

thanks for the taking the time to reply.

We could resolve our confusion. 

Best,

murb

Maybe this can help somebody else:
Turns out that when the global variable nbs is loaded in l. 146 of NBSrun.m, it would load our previous test statistics. Because we did not specifiy new contrast paths for our post hoc testing, but rather overwrote the previous files, NBSrun assumed that the input hadn't changed (it is checking for that in l. 271 ff.), and did not re-compute the edge statistics. Thus, all t-contrasts would yield the exact same results although it appeared as if NBSrun was recomputing, because the permutations were re-done.

Adding 'clear global nbs' before NBSrun solves this.