Thanks for pointing out this confusing message. This should be improved. I am in fact not sure why it was added (over 12 years ago) but the likely reason is that eigenvector centrality is not well-defined for all inputs, and directed graphs can be problematic. Generally, people use hub and authority scores in directed graphs.
Eigenvector centrality is usually justified through the Perron-Frobenius theorem. In short, this guarantees that there is a unique non-negative solution to the problem if the graph is strongly connected and all weights are positive. Most directed networks you will come across are not strongly connected, and eigenvector centrality may not be meaningful. For example, there may be more than one solution to the problem, but igraph will only return one (without warning) or may fail entirely.
Eigenvector centrality scores correspond to the values of the first eigenvector of the graph adjacency matrix.
This is well defined but only unique if the graph is strongly connected as in this example (from Edmund Landau in Über Preisverteilung bei Spielturnieren) as you pointed out.
igraph computes the left eigenvalue, i.e. the score of a vertex is proportional to the sum of the scores of vertices pointing to it (not the vertices it points to). With eigen you computed the right eigenvalue.