Ahhh I think there's a second problem here that I missed earlier: the train recipe also adds all labels present in the data to the model, which makes sense for all other scenarios – except the one where you have coarse-grained part-of-speech tags. What happens if you comment out those lines (line 101-102 in prodigy/recipes/train.py)?
Alternatively, you could also just use the previous pos.batch-train recipe. It's still included with Prodigy and the plan it to replace it with the new train recipe. But since this one use case isn't fully covered yet, there's nothing wrong with using the old recipe 