Read the qform ParaVision actually writes, and run the NIfTI oracle in CI (#182)

[gdevenyi]

Fixes #182.

test_affine_agrees_with_the_paravision_nifti_sform tripped its own assert sform_code > 0 before it ever compared a matrix, and skipped entirely in CI — so it reported nothing either way.

Read the form the writer populates

NIfTI-1 stores the affine two independent ways and a file need only carry one. All 14 PV360_StdData exports write the quaternion (qform_code=1) and leave the sform zeroed — method 2, squarely within the standard — while the helper read offset 280 regardless.

Decoding the qform instead, the export agrees with the VisuCore-derived affine to 4.8e-07 everywhere: obliques, the negative-determinant UTE3D volume and the diffusion series included. The oracle was sound the whole time and silently confirming the geometry.

nifti_affine now takes the sform when there is one, the qform otherwise, and fails only if a file declares neither — so it keeps working if a future ParaVision writes both.

Cover what is on disk

Coverage follows the nifti/ directories that exist rather than a hardcoded five, which brings in the four diffusion reconstructions plus T2map_MSME, T2star_FID_EPI and T2star_map_MGE: 14 reconstructions, 166 volumes. Every volume of an export is checked instead of whichever one glob happened to yield first, which also pins each export as internally consistent (they are — spread 0.0 within every directory).

Where no corpus is present the parameter set is empty and the test skips cleanly, rather than erroring on next() of an empty glob as it did before.

Make it run in CI

conftest smudged only 2dseq, traj and rawdata.job*, so the .nii files stayed LFS pointers and nifti_affine took its skip branch. .nii is now required media — 54 MB of a 1.7 GB dataset.

The cache key gains a -nifti suffix, because actions/cache does not re-save on a hit: without the bump the newly required files would be re-downloaded on every run instead of once. Drop that hunk if you would rather pay the download than invalidate the cache — the rest of the change works either way.

Verification

• 14/14 pass locally, over 166 .nii files.
• Negative control: the assertion rejects a missing RAS flip, a 0.1 mm origin shift, and a slice-direction sign flip. It is discriminating, not vacuous.
• Suite: 2124 passed, 12 skipped, no failures — the 5 long-standing red cases are gone and 9 net new checks replace them.

🤖 Generated with Claude Code

https://claude.ai/code/session_01NuK1cZi8U54WXAdXMmGpzy

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.