I just found an unfortunate complication for the comparison of the effective sample size.
It seems that nutpie and stan differ systematically in how they choose the step size.
Both should be using a target acceptance rate (or adapt_delta) of 0.8, but the actual average acceptance rate (ie trace.sample_stats.acceptance_rate.mean("draw")) looks like this for 1000 warmup steps:
@Bob_Carpenter or @avehtari: Any idea why stan would have a systematically higher than 0.8 acceptance rate?