For the same input across either multiple runs or one long always-on
run, we now produce the same output. This wasn't the case because our
indexing did not take into account stuff processed in previous runs.