[antonsoo]

The final bin is widened to include its right endpoint, but density uses that widened width while returning the original edges. For narrow bins at a large offset, the resulting density does not integrate to one:

import numpy as np
import boost_histogram as bh

edges = np.array([1e16, 1e16 + 2, 1e16 + 4])
density, returned_edges = bh.numpy.histogram(edges, bins=edges, density=True)
print(np.sum(density * np.diff(returned_edges)))
# Before: 0.6666666666666666; after: 1.0

Compute density bin volumes from the returned edges, using sparse meshgrids for multiple axes. The count/object return paths keep their existing behavior.

Sixteen of the 20 new regressions fail on current develop; all pass after, covering one to three dimensions, positive/negative offsets, weights and the 1D/2D entry points. On 2,048 previously recorded clock samples, the affected final-bin relative error reaches 50% before; all 12 configurations agree with independent count/bin-volume normalization to at most 2.2e-16 after. All count results and 12 object controls remain exact.

Local native suites: Python3.12, 1,238 passed/4 xfailed; Python3.14, 1,237 passed/1 skipped/4 xfailed. Full prek run -a, including strict mypy, passes.

[antonsoo]

The PyPy3.11 job fails during dependency installation: NumPy has no usable wheel for that interpreter, and the workflow disables building NumPy from source. Configure, build and tests are skipped in that job. The reported CPython native and downstream Hist checks pass; the free-threaded macOS wheel job is still running. Job: https://github.com/scikit-hep/boost-histogram/actions/runs/38534673422/job/115799656501