Just to confirm to myself that I wasn't f'ing up, I re-ran the queries from the search machine, adding remote nodes to it one at a time, and sure enough, the query time increases with the equivalent search time of each node.
If I execute the query using the cmd-line on all the remote machines (via pdsh) in parallel, then each one bounces back with the expected .1-.3 times, nice and zippy. Do 'em all together via LucyX::Remote::SearchClient, and the query time drops to sum(time*N).
