[antonsoo]

Integer starting coordinates silently truncate accepted continuous proposals while retaining their floating-point log probabilities. The stored positions and probabilities can then disagree:

import emcee
import numpy as np

sampler = emcee.EnsembleSampler(8, 1, lambda p: -0.5 * np.dot(p, p))
sampler.random_state = np.random.RandomState(117).get_state()
sampler.run_mcmc(np.arange(-4, 4)[:, None], 10)
np.testing.assert_allclose(
    sampler.get_log_prob(),
    -0.5 * np.sum(sampler.get_chain() ** 2, axis=-1),
)

The maximum discrepancy is about 2.52 on main and zero with this change. Promote integer or boolean starting coordinates to floating point before sampling. Existing floating-point types, including long double, keep their dtype. Regression cases compare signed/unsigned integer arrays and State inputs with equivalent float starts using stretch and Gaussian moves. This supports continuous sampling from integer-valued starting positions; it does not add integer-valued proposals.

Validation on Linux/Python 3.14.7: all 142 tests and repository-wide pre-commit hooks pass, including the existing long-double tests. For Bayesian regression on 107 NIST ultrasonic calibration measurements, all 108 signed/unsigned integer-start and State configurations now match their float-start chains exactly. The floating controls are unchanged; an independently derived Gaussian posterior checks every stored log probability. Other platforms were not run locally.