C++ bindings misread strided NumPy arrays #4

Several C++ bindings read NumPy buffers as though their elements are contiguous. However, py::array_t can preserve the strides of a sliced or reversed array. This can produce incorrect results without an error; a negative-stride view can also cause a read before the start of the buffer.

Reproduction

import numpy as np
from scipy import sparse
from actionet import _core

counts = sparse.random(30, 40, density=0.4, format="csc", random_state=1) * 10
labels = np.random.default_rng(0).integers(1, 4, counts.shape[0]).astype(np.int32)

storage = np.full(labels.size * 2, 9999, dtype=np.int32)
storage[::2] = labels
view = storage[::2]

# The arrays contain the same labels.
np.testing.assert_array_equal(view, labels)

expected = _core.compute_feature_specificity_sparse(counts, labels, 1)
actual = _core.compute_feature_specificity_sparse(counts, view, 1)
for key in expected:
    np.testing.assert_array_equal(np.asarray(actual[key]), np.asarray(expected[key]))

On the original build, the view can be interpreted as containing 9999 among the labels, producing different results from the contiguous array. Reversed label views can raise a RuntimeError. The same packed-buffer assumption occurs in other binding paths, including singular values passed to reduce_kernel.

Expected behavior

A NumPy view should produce the same result as a contiguous array containing the same values. Bindings that read raw pointers should account for array strides or obtain a contiguous representation first.