quasiraw: fix N4BiasFieldCorrection "not same physical space" crash

[Duplums]

Problem

brainprep_quasiraw() passes brainmask()'s output straight to ANTs' N4BiasFieldCorrection as the -x mask. mri_synthstrip (FreeSurfer) and fslreorient2std (FSL) each write their own NIfTI header independently, so the mask can end up on a grid that's numerically equivalent to the reoriented image but not bit-for-bit identical — down to the qform and sform of a single file sometimes disagreeing with each other by a float32 ULP or two. N4's ITK reader enforces exact geometry equality between -i and -x and aborts otherwise with:

itk::ExceptionObject: Inputs do not occupy the same physical space!

Running the current main quasiraw workflow against real data (OpenMind, ~72k scans across T1w/T2w/FLAIR) hit this on a meaningful share of scans, heavily concentrated in a handful of source datasets whose particular header conventions land on the wrong side of that tolerance.

Fix

Adds interfaces.regridmask(), inserted between brainmask() and biasfield(). It:

1. Rewrites the mask onto the reference image's exact voxel grid (via nibabel, same shape/data, just re-anchored).
2. Transcribes the raw qform/sform NIfTI header fields (quaternion components, offsets, sform rows, codes) byte-for-byte from the reference image's own header.

Step 2 matters on its own: I initially tried flirt -applyxfm with an identity transform, and separately nibabel's set_qform()/set_sform() with the reference's affine — both still failed on one case, because NIfTI's qform is quaternion-based, and any fresh matrix → quaternion decomposition (including nibabel's own) can round differently than whatever decomposition originally produced the reference's qform (e.g. FSL's). The only way to get a genuinely identical header is to not recompute anything and copy the raw fields directly.

Testing

Verified against real OpenMind scans on the actual container pipeline (bind-mounting this branch's source over the installed package, no rebuild needed):

• 30/30 previously-failing scans (sampled across 3 source datasets, both FLAIR and T1w) now succeed.
• 5/5 already-passing scans still succeed (no regression).
• Spot-checked output geometry/intensity stats for sanity on several outputs.

🤖 Generated with Claude Code