I'd certainly be surprised to have the order of zero and negative values in the input have such a dramatic effect on the result. It would make debugging these cases unnecessarily confusing and difficult. And I think it is indeed worth fixing, as these edge-cases are likely to come up of time when dealing with large amounts of data.

I'm +1 on Warren's latest suggestion for a fix.

P.S If we agree on this, either Warren should make the fix himself or we should mark it "newcomer friendly" and let a newcomer handle it.