I think kmod fixed some of this cases, but I guess nobody really looked into where where the additional refs come from when running the numpy tests. So not sure how hard it is to fix it. But in general I think being more conservative / having more references is not really something I see as a big problem. It's unfortunate that we miss out on some refcnt == 1 optimizations and with numpy there are some user visible problems with it. But still I think it's a smaller issue and something which one has to expect when switching between different python implementations.