The distinction between nominal/ordinal in the last version was not required because discrete models generally aren't classified this way (the imposed scoring coefs had no effect on the model fit/log-likelihood). So items are all technical 'nominal', and just represent the probability that a given class (or combination of classes) answer a given category. See the dev version for other kinds of discrete models with mdirt (e.g., DINO/DINA, multidimensional latent classes, etc), which also has better support for things like M2, itemfit, fscores, and so on.