Yes, this what we have often seen in other problems. With lighter-tailed priors that allow those points but make them less likely, it is still possible that likelihood and prior are in conflict. You will then not see cliff edges, but you should compare prior and posterior and check the sensitivity to prior changes. Last time Balancing loo performance and Prior sensitivity analysis the similar problem with prior vs loo, the reason was a typo. You should also check the loo diagnostics, as failing importance sampling can also sometimes describe this kind of behavior.