This has been a longstanding problem in neuroimaging (https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0007200).
The problem is that Harvard atlases were produced with the MNI152 Linear version of the MNI templates. Although the MNI152Nlin2009cAsym is well aligned with the linear version, there will always be inconsistencies between atlases if they would have defined using the MNI152Nlin2009cAsym template as well.
In summary, by applying an atlas defined w.r.t. the Linear version of MNI152 on data aligned to MNI152Nlin2009cAsym you are dismissing some accuracy errors derived from the differences between templates.
Then you would think, why didn’t you use MNI152 linear as the registration target in the first place? The answer is that then you would be introducing similar accuracy errors but on your registration step. We considered that it was more important to have the best alignment possible to the standard space, assuming that the current lack of atlases defined on MNI152Nlin2009cAsym would be addressed at some point in time.
For these reasons, we are working on an easy way of transferring atlas information through different templates. That would be one of the objectives of the project I mentioned above (templateflow).
Hopefully, this addresses your concerns