The rules toward variance in Gaussian node produce improper InverseGamma distributions that is incompatible with projection like methods

[bvdmitri]

I guess, this is intentional, but the rules defined in normal_mean_variance/var.jl produce GammaInverse with α explicitly set to -1/2. This renders the distribution as improper and does not allow to use it with our available approximation methods.