You could use also use az.plot_pair with divergences=True instead of corner to see if there are any regions/boundaries/pairwise combinations that are causing you divergences.
I’ll note for the record that the paper you link didn’t include any sampler diagnostics, which is a bit sus (as the kids say). The model might just be really hard to sample, and re-parameterization is indeed needed.
Looking more closely at the priors and combinated with Ricardo’s comment, it might help to replace those uniform priors with something that would help pin down one of the posterior modes. You might not have any reason a priori to believe one over the other though…