I screwed up my local test. I'm getting 'error' here too.

It's correct by the definition of 3-way merge. As far as 3-way merge can see
(and thus any tool you can use with hg), the result 'error' is correct.

And I also consider it the correct behavior for hg. The alternatives are:

a) use a significantly better automated merge algorithm 

b) forego automatic merging

I consider the pain caused by (b) to be MUCH worse than the pain caused by the
relatively rare case of incorrect automated merges. And basically every SCM
since at least RCS has made the same choice.

As for (a), it's not clear that a suitable algorithm exists.