do your chains have the same length?
bpcomp computes the max-diff statistic based on the same number of points for each chain. Which means that, if the chains have different lengths, then all will be read until the size of the shortest one has been reached.
and I can imagine cases where the maxdiff is small on the first part of the 3 chains, but then increases again.
in any case, don't hesitate to send me the files, if you want me to check what happens here