Resample lesion mask into in_file's voxel grid

[tsalo]

Closes #1021.

Summary

create_cfm() builds the cost-function mask used to constrain registration by subtracting a lesion mask from a global/brain mask. It loaded the lesion via nb.as_closest_canonical(), reorienting it to RAS, while leaving in_file in its stored orientation. When in_file is not RAS, the two voxel arrays are in different orders, so the lesion gets subtracted from the wrong voxels — the lesion has no effect at its true location and spurious differences appear elsewhere.

This reorients/resamples the lesion mask into in_file's exact voxel grid with nibabel.processing.resample_from_to(..., order=0) before subtracting. Nearest-neighbor keeps the mask binary, and resampling by world coordinates makes the result correct for any orientation or grid — not just RAS.

Reported downstream in PennLINC/qsiprep#1023, where the normalization runs in LPS and every lesion ROI hit this bug.

Changes

• niworkflows/interfaces/norm.py: replace as_closest_canonical lesion handling in create_cfm with resample_from_to(..., order=0); update the now-inaccurate docstring note that claimed the inputs must share a grid.
• niworkflows/interfaces/tests/test_norm.py: add test_create_cfm_lesion_orientation, which builds an LPS in_file and an RAS single-voxel lesion at a known world coordinate and asserts the cost-function mask is excluded at exactly that world location. Fails on the old code (excluded voxel lands at the mirrored location), passes with the fix.

Testing

pytest niworkflows/interfaces/tests/test_norm.py — passes (incl. the new regression test); no other tests affected.

🤖 Generated with Claude Code