Pattern files for textcat.teach

Yes, your solution is really elegant, actually! Since the annotations all have the same format and are collected with the same process (binary feedback on text plus label), you could store everything in one dataset, too. It's just make it more difficult to revert your changes if you make a mistake or want to try something else in the textcat.teach step.

The --patterns approach on textcat.teach was intended to be the "simpler solution" – but as it turned out in this thread, it does have some limitations, especially for getting over the cold start problem with rare categories.

Yes, for the final model, you should be able to train on all the annotations from scratch. At least, there is no reason why you should have to use your pre-trained model as a baseline. If you want, you could try both approaches and compare the results – if there's a significant difference, this would be very interesting!