test(3d ngldm): demote the family's one vetted row, and pin its wrong values honestly

[darkclad]

The last family in the per-family series, and the only one that closes its gap row by demoting it. 3NGLDM_DCP read status=vetted, oracle=mirp from an offline harness run with target_test naming test_3d_ngldm_mirp.h, a file that does not exist. Running that oracle shows the claim should never have been made - and shows why the other 18 rows are right to be regression.

The oracle run

gen_ngldm3d_mirp.py runs MIRP 2.6.0 on the segmented phantom at a config-matched recipe: fixed_bin_number n=64 against Nyxus' GREYDEPTH=64, distance 1, alpha 0, by_slice=False. D3_NGLDM_feature bins on STNGS_NGREYS, the generic grey depth, so there is no binning mismatch to explain the results. MIRP disagrees on 16 of the 17 features it can compute:

GLNU 115443 vs 4350.27 26.5x
DCENE 0.143484 vs 0.00287482 49.9x
GLNUN 0.225757 vs 0.0158519 14.2x
HDE 261.018 vs 28.0738 9.3x
DCV 86.1706 vs 11.9476 7.21x
HDLGLE 0.0252015 vs 8.71408 0.00289x
LGLCE 0.000359684 vs 0.0321849 0.0112x
... full table in the audit report

3NGLDM_GLM and 3NGLDM_DCM have no MIRP counterpart at all (no gl_mean / dc_mean column).

Two causes, both in the source

calc_ngld_matrix iterates the ROI bounding box and explicitly declines to skip background ("// Do not skip off-ROI pixels", with the guard commented out). On this phantom that is 551,040 voxels against an ROI of 274,432 - a factor of 2.008. IBSI NGLDM is defined over ROI voxels only. This predicts the error pattern rather than merely correlating with it: background voxels bin alike and each sees ~24 identical neighbours, so they pile into one grey row at maximum dependence, which inflates a sum-of-squares over grey rows superlinearly (GLNU), raises energy (DCENE), lowers entropy (DCENT) and pushes the dependence distribution to its top end (HDE up, LDE down). DCNU, which scales closer to linearly with voxel count, sits at 2.09x - the bounding-box ratio.

Second, the shifts table holds 24 entries where a 3D Chebyshev-1 neighbourhood has 26: the two pure-axial neighbours (0,0,+1) and (0,0,-1), directly above and below the centre, are absent. nsh is therefore 24, and maxNr = nsh + 1 carries a comment reading "max dependence 8 (due to 8 neighbors)", which is the 2D count.

Neither is fixed here - both change 19 public feature values and belong on their own branch.

Why DCP is demoted rather than promoted

It is the family's only agreement and it agrees at 1.0 on both sides. Dependence-count percentage is the fraction of voxels with at least one dependency; it is exactly 1 on this phantom under both the correct and the incorrect neighbourhood, with or without background voxels. It is an assertion that cannot fail for the reasons we would want it to fail. Promoting it would leave the registry saying 3D NGLDM had been checked, on the strength of a degenerate constant standing beside sixteen features that miss the same tool by up to 50x. The row now reads status=regression, candidate_oracle=mirp, flag=implementation-defect.

That leaves the family with zero vetted rows, the only one in the series. It was zero before as well; the difference is that the registry now says so.

The pins are wrong on purpose

All 19 goldens are regenerated at full precision and the band tightened from 10% to rel=1e-9. The old pins were two- and three-significant-figure numbers (0.1, 261, 740, 0.00036) inside a +-10% band, which cannot detect the very fix these values are waiting for. A drift guard's job is to notice change, and these will change; test_3d_ngldm_dump_regression() regenerates them when the implementation is corrected, at which point the MIRP comparison should be re-run to promote the family properly.

Test file cleanup

test_3d_ngldm_regression.h carried a dead #if 0 block (a superseded copy of the assert body) and relied transitively on , , , , and helpers/fsystem.h for fs::exists; all now direct. Its golden lookup used operator[], which default-inserts a missing key as 0 and then compares against that fabricated reference - now guarded with find(). The 2026-07 MIRP comparison that lived inline in the header moves to the audit report, at full precision and with causes attached.

For the record, since an earlier note in this series said otherwise: the 19 tests in this file DO assert. Verified by negative control - perturbing 3NGLDM_DCENE from 0.14 to 0.99 makes TEST_3D_NGLDM_DCENE_REGRESSION fail at the assertion.

Registry

All 19 rows get a config_recipe id (two new: ngldm3d.regression_ut_phantom and ngldm3d.mirp_fbn64 - the second is referenced but not asserted against, so the divergence stays reproducible and the promotion can be re-run), a real tolerance, target_test cleared, source moved off tracker, and current_test synced to the files that cover each feature.

Verified: runAllTests 774/774 on Windows and under ASan+UBSan on Linux; pytest tests/python/ full directory; check_coverage.py --check, check_test_names.py --check and audit/scan_ngldm3d_coverage.py --check all clean.