The stored transformed training data, learner.Xt_training, needs to be updated every time the model is re-trained. Otherwise, if a transformation starts adding more/less feature columns, the old and newly-transformed data instances cannot be concatenated.

See Nawab Hussain's comment