What would you think about letting it go until it hits max_treedepth
before evaluating whether Delta H > 1000
?
Then we’d end up wasting substantial resources every time a trajectory diverged.
If there are hundreds of divergences, yeah, but then it the results are irrelevant. If you have 5 potential divergences, letting them go a few more leapfrogs steps to see if they turn around is not going to waste substantial resources.