[bpolverini]

You're right. To be honest, the reason why we don't use it the way it is described in the paper, is entirely out of fear of misconfiguring the Skein context. The NIST reference implementation for Skein doesn't provide a simple way or an example for building the valid configuration for the HMAC construction they recommend, and we always wanted to feel like we had a wall to our back (i.e. being able to check against the NIST implementation or to check against test vectors). They do mention in the paper that doing the keyed mac approach is provably safe, which just felt easier to implement.