The current admission control treats all backends symmetrically and sets aside the mem_limit. This makes sense for now given that we have the same mem_limit setting for all backends.
One case where this could be somewhat problematic is if you have dedicated coordinators with less memory than the executors, because the coordinator's process memory limit will be fully admitted before the executors.
If you have multiple coordinators and queries are distributed between them this is relatively unlikely to become a problem. If you have a single coordinator this is more of an issue.