The client can go away and come back later with a different set of NIDs. With the server locking client NIDs, it is possible that NID which used to belong to the client and got locked as primary on the server now belongs to a different client. The server will attempt to merge peer records, but may be unable to do it properly because deleting of the primary NID is not allowed.
Let square brackets designate actual node configuration and round brackets - LNet view of the peer. For example, suppose the client with NIDs [A,B] connects to the server. The server creates peer record (pA,B) where NID A is primary. Then the same client goes away and comes back only with NID B configured. This results in "[B] - (pA,B)" because the server won't delete the locked NID A for the client. If another client comes up with NID A configured, there will be a conflict because the server will delete NID B from the existing peer record, so the original peer is kind of hijacked for the new client: "[A] - (pA), [B] - ?".
Not sure exactly what this may cause from the FS point of view, but it looks wrong. The 54539 patch is trying to avoid such confusion by making sure that the server discovers the client NIDs before using them.
Merged for 2.16