I'm still a little confused and will have to go over some code or something to really understand the limitations of what transducers can do... can any transducer be used in a "parallel" context (like map and filter) or are they limited to a linear context (like the fold makes me suspect)?
I think it could be parallelized by chunking up the input, however the transducer doesn't know how to do that by itself, so the step function would have to be modified somehow.
Check out clojure reducers (http://clojure.org/reducers) they are earlier work in the same line of thinking. The where used to do pmap style things. The same tricks can be used with trnasducers, but I dont think its implmented yet.