A shiny new quarter that your model isn’t identified in a subtle direction, so that the resonances look identified but the posterior itself is pathological and the dynamic HMC trajectories are forced to grow too large. RHMC may help a bit in this case, but any improvements shouldn’t be drastic.
A possibly related issue is that, because the rotation group has a different topology than the reals, any model with a completely free rotation will be difficult to fit in Stan directly (regardless of how the rotation is parameterized).