test(3d morphology): assert the five PCA axis features against MIRP, and two MATLAB rows nothing checked

[darkclad]

Closes the five 3D morphology rows that read status=vetted, oracle=mirp with no in-tree oracle assertion, plus two more the family plan did not count.

The five axis features

3MAJOR_AXIS_LEN, 3MINOR_AXIS_LEN, 3LEAST_AXIS_LEN, 3ELONGATION and 3FLATNESS carried a mirp verdict from an offline harness run (source=tracker) with target_test naming a _regression file. Nothing in the tree compared them against MIRP; the only thing touching them was the coverage sweep, which pins Nyxus' own output as a snapshot.

gen_morphology3d_mirp.py now runs MIRP 2.6.0 against the segmented phantom and test_3d_morphology_mirp.h asserts all five. Same definition on both sides - 4*sqrt of the mask covariance eigenvalues, and their ratios - so the agreement is exact to double precision:

3MAJOR_AXIS_LEN 104.70681271508683 rel 0
3MINOR_AXIS_LEN 88.30145986864228 rel 0
3LEAST_AXIS_LEN 71.51449974198198 rel 2.0e-16
3ELONGATION 0.8433210559938976 rel 0
3FLATNESS 0.6829975804590384 rel 1.6e-16

Asserted at rel=1e-9, far tighter than SPEC 7's rel=1e-3 same-definition tier, because that is what the measurement supports.

The generator checks the identities the old defect broke -------------------------------------------------------- This family carried an eigenvalue-ordering defect that produced 3LEAST_AXIS_LEN > 3MAJOR_AXIS_LEN and 3FLATNESS > 1, both structurally impossible. It is fixed in 3d_surface.cpp and these are the corrected values. The generator now asserts, on MIRP's own output, that MAJOR >= MINOR >= LEAST > 0, that ELONGATION and FLATNESS lie in [0,1], and that each equals its defining ratio - so a misconfigured oracle run is caught before anything is pinned. No _invariant file is added: the five oracle pins fix the exact values on this fixture and are strictly stronger than the bounds would be there.

Two rows the plan did not count

3VOXEL_VOLUME and 3VOLUME_CONVEXHULL read status=vetted, oracle=matlab and had a MATLAB golden and a stated band in morphology_3d_matlab_ref_vals/_ref_tols - but no oracle-named test asserted them. test_3d_morphology_matlab.h registered exactly one function, for 3MESH_VOLUME. The only thing comparing the other two against MATLAB was the parameterized sweep in test_3d_morphology_coverage.h, whose case names carry no oracle token, so the claim rested on a test that does not say what it checks (SPEC 6.2). The report generator misses this the same way it missed two rows in 2D morphology: a feature name in an oracle file counts as coverage whether an assertion reads it or only a table does. Both now have an assertion through the existing helper, goldens and bands. Both pass.

MATLAB cannot be re-run, so MIRP cross-checks it

Three rows claim oracle=matlab and revet step 3 says to run every claimed oracle. Not possible here - no licence, and Octave's image package has no regionprops3 - and the goldens have no in-repo generator (not_covered.md C). The generator prints the corresponding MIRP quantities instead, unpinned:

voxel volume MIRP 274432.0 MATLAB 274432.0 exact
convex-hull volume MIRP 496958.32 MATLAB 497824.0 0.17%

Both MATLAB goldens are independently corroborated, which also relocates the disagreement: Nyxus' 3VOLUME_CONVEXHULL of 478516 is 3.6% from MATLAB and 3.7% from MIRP while the two tools sit 0.17% apart, so the difference is Nyxus' discrete voxel hull against two triangulated ones - which is what the existing 5% band documents, now with a second measurement behind it.

The surface-area convention gap stays open

3AREA counts exposed voxel faces (59992) where MIRP integrates a marching-cubes mesh (46739), a 28% convention difference that no tolerance turns into an agreement. 3AREA and the five features derived from it stay status=regression, with the reason recorded per row rather than left as a bare absence. Settling it changes six public feature values and belongs on its own branch.

Test file cleanup

test_3d_morphology_common.h carried a dead #if 0 block - a superseded copy of the whole fixture, ending in a line of asterisks that is not valid C++ and only compiled because it was disabled. Removed. It also included test_ref_vals.h while declaring no reference table. Include hygiene across the family: _common.h relied transitively on , , and helpers/fsystem.h; _matlab.h named Pixel3 and Nyxus::calc_eigvals without including features/pixel.h or helpers/helpers.h, and used std::abs/std::vector without /; _regression.h relied on . All now direct. Three headers carried "Migrated from test_3d_shape.h (Wave 8)" history, which belongs in the audit report.

Registry

All 14 rows get a real config_recipe id (three new: morphology3d.mirp_ibsi, morphology3d.matlab_regionprops3, morphology3d.regression_ut_phantom - one fixture, three benchmarks, so unusually they ARE comparable to each other, which is what makes the MIRP cross-check meaningful), a real tolerance, target_test cleared, source moved off tracker, and current_test synced to the files that actually cover each feature.

Verified: runAllTests 780/780 on Windows and under ASan+UBSan on Linux; pytest tests/python/ full directory (92 passed, 1 skipped); check_coverage.py --check, check_test_names.py --check and audit/scan_morphology3d_coverage.py --check all clean; gen_morphology3d_mirp.py re-verifies all five goldens and all four structural identities against a fresh MIRP 2.6.0 run.