There's nothing special here. If you have a subpopulation with fewer than
buildSNNGraph() will forcibly construct edges between cells in that subpopulation and cells in other subpopulations. This increases the risk that the subpopulation will not form its own cluster as it is more interconnected with the rest of the cells in the dataset.
I guess the wording of the documentation is misleading, as the interpretation of
k is that of the anticipated size of the smallest subpopulation. It is not a specification of the size of the smallest cluster that you are willing to obtain. The actual minimum cluster size is at the mercy of the community detection algorithm that you choose, if it enforces (explicitly or otherwise) a minimum cluster size at all.