[proost]

ensure_enough_sections() doubled the number of sections even when the
section size was already MIN_K and the reduced size rounded back to
MIN_K. The nominal capacity then grew 2x instead of the sqrt(2).

This is same as java.

[leerho]

See comment in main thread.

[leerho]

Thanks, Hyeonho. The change looks right to me: it adds the same sectionSize > MIN_K stop that ReqCompactor.ensureEnoughSections() has in Java, and the new test fails without it (k = 14 gives a nominal capacity of 384 instead of 192).

One small request: please restore the commented-out "for manual comparison with Java" test, or move its removal to a separate change, since it isn't related to this fix.

We're finishing a bug-fix release and holding other PRs until it's out, so this will wait until then. It only affects a few small values of k, and the sketch is larger than designed rather than wrong, so it doesn't need to go into that release. It may need a small test adjustment at merge time, depending on what lands on master first.