test(vetting): verify 2D first-order with MATLAB R2026a

[vjaganat90]

Why this is needed

The 2D first-order family still carried a MATLAB correctness claim without a checked-in MATLAB generator. The registry claimed 33 oracle=matlab features, while test_2d_firstorder_matlab.h contained 28 literals with no tool version, exact functions, recipe, benchmark, or runnable provenance. Those two sets did not even agree: six registry claims had no MATLAB assertion, while the header called histogram UNIFORMITY a MATLAB assertion even though the registry correctly treated PyRadiomics as its oracle.

The old audit had already found the central problem: several percentile literals were Nyxus outputs, not values produced by MATLAB or Octave. This PR replaces that inherited claim with a licensed MATLAB R2026a run and lets the measured results decide the disposition feature by feature.

What the licensed run establishes

gen_firstorder2d_matlab.m downloads the canonical pixel fixture and golden header from the moving PolusAI/nyxus main tree. It extracts the 154 fixture intensities directly from tests/test_data.h, invokes native MATLAB statistics built-ins, checks the generated and pinned feature sets in both directions, prints full-precision values side by side, and fails if a pin exceeds its declared band.

MATLAB honestly covers 31 features:

• Direct built-ins cover extrema, mean/median/mode, sample and population variance/stddev, skewness, kurtosis, MAD, percentiles, IQR, RMS, energy, and integrated intensity.
• Derived values use only the feature-defining arithmetic over those built-ins: COV, excess kurtosis, fifth/sixth standardized moments, QCOD, standard error, PIU, and slide-relative intensity range.
• No Nyxus histogram or first-order algorithm is reproduced.

The C++ table now pins the exact MATLAB R2026a values. For the eight percentile-derived features, Nyxus uses its fixed 100-bin interpolated CDF while MATLAB evaluates the raw sample. They still agree within the measured rel=3e-2 band; the worst residual is QCOD at 2.75%. The other 23 assertions use the SPEC same-definition tier, rel=1e-3.

This restores honest MATLAB coverage for P01, P25, P75, P99, and QCOD, and replaces the old Nyxus-derived pins for P10, P90, and INTERQUARTILE_RANGE with the actual MATLAB values.

What remains outside MATLAB

This PR does not force incompatible semantics into the MATLAB file:

• UNIFORMITY remains PyRadiomics-vetted. Producing it here would require reconstructing Nyxus histogram discretization.
• MEDIAN_ABSOLUTE_DEVIATION is moved to regression: MATLAB mad(x,1) is the median absolute deviation, while Nyxus computes the mean absolute deviation about the median.
• ROBUST_MEAN remains regression: MATLAB rank trimming and Nyxus histogram-derived P10/P90 selection are different definitions.
• ENTROPY and ROBUST_MEAN_ABSOLUTE_DEVIATION remain PyRadiomics-vetted. The redundant standalone PyRadiomics assertion for the latter is removed; the table-driven oracle assertion already covers it. (I can revert this redundant code elimination but I think we should remove it here)

The honest family result is therefore 34/36 features vetted by at least one oracle, with the two unresolved semantics explicitly regression-only. The generated global report drops two stale vetting claims instead of preserving an inflated headline.

SPEC alignment

• Adds the licensed MATLAB R2026a generator and complete golden-site provenance.
• Consolidates the 31 MATLAB comparisons into one table-driven gtest with per-feature tolerance selection.
• Adds firstorder2d.md, enumerating the MATLAB, PyRadiomics, IBSI, and two production-only regression config points.
• Defines the canonical 154-pixel benchmark and every referenced config recipe.
• Rewrites the family registry as 64 assertion rows: 31 MATLAB, 18 PyRadiomics, 12 IBSI, and 3 regression. Every row names its exact recipe, tolerance, test, gtest case, benchmark, and in-tree source.
• Replaces the stale Octave regeneration guidance and audit language with the licensed run, exact function mapping, measured values, and honest exclusions.

MATLAB remains an offline golden generator; CI consumes only the checked-in C++ literals.