Instead of merging the MMFs with an external script, would it work
in principle to:
- first output the trees but not the clustered models
(i.e., launch HHEd with the TB and ST commands but no -w switch);
- then, once the trees are built, use an HHEd run that would
consider the list of full-context models as unknown models
to be built by applying the trees to the initial collection
of states, thus effectively selecting the state cluster
representatives and "forgetting" the other states?
(I.e., load the initial models with -H, then load the trees
with a series of LT commands, then use an AU command over
the initial list of full-context model names, but this time
with a -w switch to output the clustered models in place of
the original un-clustered ones.)