This patch seems to resolve the issue by propagating the number of reducers argument through to the back-end processing steps where the actual output vectors are produced. It also includes a slight modification to SequenceFilesFromDirectory to remove chunk-size upsizing to 64mb which allows Reuters data to be split into 3 smaller files to improve processing. All unit tests run.
The attached build-reuters.sh runs LDA iterations in about 1.5 min vs. 5.5 min with a single vector file on a 3-node cluster using 3 mappers and 2-3 reducers for the vectorization. I will commit it in a day or so but want some more eyeballs on it since this is new code for me.