Looking at the code now, I was unable to find a good reason as why we only rely
on mergestate extras info after checking whether a filelog parent is ancestor of
other or not.
I mean if we have stored in mergestate that other was chosed, we should
blindly pick that one.
This cleanup will also help introduce more cases when both fparent1 and
fparent2 are non-null but using info from mergestate, we can fastpath.
The test change actually demonstrates the point of the patch. During merge we
were getting the other side of the file but on commit we were marking that as
merged.