But the more I think about it, this shouldn't affect the sample accuracy in any way. So forget about this question.
OK, but it was easy so I did it anyway as part of a test in a new project.
I cannot reproduce my results in the new project, so I'm going to give up; the problem is in a project that is very complex and has been through many convolutions over the years; I'll probably never figure the source of the problem if it's specific to that project. So I'll just recheck those sections with my ears.
For the record, on the new project I did the following:
1. Opened a different template.
2. Imported a stereo background mix and a mono vocal.
3. Tested and got perfect nulls for:
a) the stereo bg
b) the stereo bg with plain vocal
c) the bg with vocal played by Melodyne (after transfering the vocal to M)
d) the bg with vocal pitch-changed and moved in time in Melodyne as it is played back.