I was being dense - turns out I just needed to use pm.Potential.
@jonsedar Thank you - I will try this when I need the pointwise log-likelihood.
@jessegrabowski For identifying the model, try setting r^*=z_t. One of the (many, imo) problems with HLW is that because there are no constraints on z_t, the model with r^*=g_t + z_t is observationally equivalent to the model with r^*=z_t. Kiley 2015 does this as well.