[antonsoo]

Stopping a sample loop early leaves reserved space in the default memory backend. Continuing with fewer steps than the unused reservation raises ValueError: negative dimensions are not allowed:

import emcee
import numpy as np

coords = np.random.RandomState(0).normal(size=(16, 2))
sampler = emcee.EnsembleSampler(16, 2, lambda p: -0.5 * np.dot(p, p))
for i, state in enumerate(sampler.sample(coords, iterations=1000)):
    if i == 9:
        break
sampler.run_mcmc(state, 20)
assert sampler.iteration == 30

Reuse existing storage until more is needed, and initialize a new blob array across the full reserved capacity. The regressions compare an early-stopped/resumed run with an uninterrupted chain, including structured blobs and thinning, and retain the HDF backend as a control.

Validation on Linux/Python 3.14.7: all 147 tests and repository-wide pre-commit hooks pass. On 107 NIST ultrasonic calibration measurements, all 72 continuation configurations match uninterrupted runs exactly, including positions, log probabilities, blobs, accepted-step counts and RNG state. This repairs all 36 failing memory-backend runs; all 36 HDF controls remain unchanged. This is continuation correctness, not a new convergence or statistical-efficiency claim.