So I made a 2nd pass (same Colab notebook) on the model and got rid of the binary latent variables by trying out a new representation for active/inactive changepoints with a sigmoid function, but not as a replacement for the > operator. No divergences now!
