We actually have built-in HMM code now, which will be a lot faster. See:
I guess we need to update the user’s guide to keep up!
Do you mean label switching? While that’s a problem for convergence monitoring, it’s usually not a problem for downstream inference, which will again marginalize out the labels.
You can do that, but it’d be more efficient to do so in generated quantities just for convergence monitoring. It doesn’t solve the underlying problem that the model’s not identified in its parameters as declared.