This paper considers the problem of learning the unknown structure of a network with the underlying topology given by a polyforest (a collection of directed trees with potentially multiple roots). The main result is an algorithm that consistently learns the network structure using only second-order statistics of the data. The methodology is robust with respect to the presence of unmeasured (latent) nodes: the algorithm detects the exact number and location of the latent nodes, when they satisfy specific degree conditions in the actual network graph. It is shown that the same degree conditions are also necessary for a consistent reconstruction. Thus, the proposed reconstruction algorithm achieves the fundamental limitations in learning the structure of a polyforest network of linear dynamic systems in the presence of latent nodes. This paper overcomes the limitations of previous results that only addressed single-rooted trees, tackling the problem in an efficient way since the computational complexity of the derived algorithm is proven to be polynomial in the number of observed nodes.