mean_accept_prob significantly different after warmup

I notice that after warmup, the mean_accept_prob significantly higher than both target_accept_prob and the mean_accept_prob observed during warmup, even on a trivial isotropic gaussian example. Minimum working example:

import jax.numpy as jnp
from jax.lax import scan
from numpyro.infer.hmc import hmc

def potential(x):
    return 0.5 * jnp.sum(x**2)

d = 10
nwarmup = 100000
nsamples = 100000

init_kernel, sample_kernel = hmc(potential, algo='HMC')
hmc_state = init_kernel(init_params=jnp.zeros(d), num_warmup=nwarmup, adapt_step_size=True, adapt_mass_matrix=False)

hmc_state = scan(lambda s, _: (sample_kernel(s), None), hmc_state, None, length=nwarmup)[0]
print("post warmup", hmc_state.mean_accept_prob)

hmc_state = scan(lambda s, _: (sample_kernel(s), None), hmc_state, None, length=nsamples)[0]
print("post samples", hmc_state.mean_accept_prob)

outputs:

post warmup 0.7992318
post samples 0.97852784

am I misusing something here?

Start from the hmc entry point and run the supplied isotropic Gaussian reproduction with the reported warmup and sampling lengths. Trace how mean_accept_prob is updated across warmup and sampling, then determine whether the discrepancy is expected or needs a regression test and correction.