Warp a volume and write selected differential features

This pipeline applies one 3-D affine transform and then requests only gradient and Laplacian outputs. Exact structured buffers make allocation ownership explicit while Affine and Differentials remain independent modules.

warped volume: (24, 28, 32) float32
gradient components: 3
Laplacian range: -63.4631462097168 73.38594055175781

import numpy as np

from splineops.affine import AffinePlan
from splineops.differentials import DifferentialPlan, DifferentialResult

shape = (24, 28, 32)
volume = np.random.default_rng(20260719).standard_normal(shape).astype(np.float32)
radians = np.radians(-6.0)
matrix = np.array(
    [
        [np.cos(radians), -np.sin(radians), 0.0],
        [np.sin(radians), np.cos(radians), 0.0],
        [0.0, 0.0, 1.0],
    ],
    dtype=np.float32,
)
center = (np.asarray(shape, dtype=np.float32) - 1.0) / 2.0
affine = AffinePlan(
    shape,
    matrix,
    center - matrix @ center,
    degree=3,
    mode="mirror",
    dtype=np.float32,
)
warped = affine(volume)

differentials = DifferentialPlan(shape, spacing=(0.8, 0.8, 1.5))
output = DifferentialResult(
    tuple(np.empty_like(warped) for _ in range(3)),
    None,
    np.empty_like(warped),
)
features = differentials(
    warped,
    gradient=True,
    hessian=False,
    laplacian=True,
    out=output,
)

print("warped volume:", warped.shape, warped.dtype)
print("gradient components:", len(features.gradient))
print(
    "Laplacian range:", float(features.laplacian.min()), float(features.laplacian.max())
)