The makeindex functions only dump the accumulated
postings when they find an end-of-file tag. This works
fine for TREC data, but HTML data we don't allow
end-of-file tags at all. Thus the postings are never
dumped, and the memory usage continues unconstrained.
I've patched 0.5.2 (will release in 0.5.3) to work
around this, but the true solution is to redo the
indexing to potentially dump postings when allocating
memory for postings. Thanks to Jason Zien
<jasonz@almaden.ibm.com> who reported this bug.