fix(glcm): default co-occurrence offset to 1 and exclude background

[darkclad]

Two production-path defects the hard-coded phantom tests missed:

• GLCM_OFFSET was zero-initialised when settings are compiled from the production path, so dx=dy=0 -> every pixel co-occurs with itself -> CONTRAST=0 for any image. Default GLCM_OFFSET=1 (IBSI delta=1) in env_features.cpp (plus GLCM_GREYDEPTH/GLCM_NUMANG).
• The MATLAB grey-binning path maps off-ROI background (0) to level 1; the old zero-skip guard only ran on the IBSI path, so background inside a concave ROI flooded the co-occurrence matrix and diluted CONTRAST. Skip on the ORIGINAL intensity (imR==0) in all binning paths (glcm.cpp).
• Soft-NAN guard CORRELATION/INFOMEAS1 for the degenerate single-grey-level case that a correctly-masked tiny ROI can produce (glcm.cpp/.h, glcm_nontriv.cpp).

Refresh the 25 background-sensitive phantom goldens in test_glcm.h (the background-insensitive keys are unchanged), add test_glcm_bugs.h with a production-path offset-default regression guard (registered in test_all.cc), and add GLCM contrast/background Python tests validated against PyRadiomics/MIRP.

[darkclad]

Justification — golden value changes for main-glcm-defaults-and-background

Branch: main-glcm-defaults-and-background (split from main-feature-validation, based on main)
Source fix: glcm.cpp, glcm.h, glcm_nontriv.cpp, env_features.cpp
Test files touched: tests/test_glcm.h (phantom goldens), tests/test_glcm_bugs.h (new),
tests/test_all.cc (registration), tests/python/test_feature_bugs.py (2 tests)

The bugs fixed

1. Co-occurrence offset defaulted to 0 (env_features.cpp). When the GLCM settings are compiled
   from the production path, GLCM_OFFSET was zero-initialised, so dx = dy = 0: every pixel
   co-occurs with itself → a purely diagonal matrix → CONTRAST = 0 for any image. The hard-coded
   C++ phantom tests set offset = 1 explicitly, so they never exercised the default and missed this.
   Fix: default GLCM_OFFSET = 1 (IBSI delta = 1), plus GLCM_GREYDEPTH/GLCM_NUMANG.
2. Background counted in the MATLAB binning path (glcm.cpp calculateCoocMatAtAngle). The
   non-IBSI (MATLAB) grey-binning maps off-ROI background (original intensity 0) to level 1, and the
   old guard only skipped zeros on the IBSI path. So background inside a concave ROI's bounding box
   was counted as a real grey tone and flooded the matrix with spurious diagonal mass (diluting
   CONTRAST). Fix: skip on the original intensity (imR.yx(...) == 0) in all binning paths, so
   every path agrees with the IBSI path and external oracles.
3. Soft-NAN guards for CORRELATION (f_corr) and INFOMEAS1 (f_info_meas_corr1): once background
   is correctly excluded, a tiny/uniform ROI can leave a single grey level → zero-variance marginal →
   0/0 = NaN. Return the soft-NAN sentinel instead.

tests/test_glcm.h golden changes (the 25 refreshed phantom keys)

These are Nyxus-convention regression values for the digital phantom (100 grey levels, offset 1,
Nyxus's asymmetric cooc matrix) — not an external oracle for the phantom itself. They were refreshed
because the phantom's slices z2–z4 contain masked-out (background) pixels, and fix (2) now excludes
those pixels, so the co-occurrence matrix — and every statistic derived from it — changes on those
slices.

• What is the ground truth: the values are the corrected code's deterministic output on the
  phantom, taken verbatim from the oracle-validated main-feature-validation. The C++ test
  (test_glcm_feature) re-passes with them (its agrees_gt(..., 100.) is a loose regression bound).
• Internal consistency: keys insensitive to the background exclusion — GLCM_CONTRAST,
  GLCM_CLUTEND, GLCM_SUMVARIANCE, GLCM_IDN, GLCM_IDMN — were kept unchanged at full
  precision. Only the 25 background-sensitive keys moved. GLCM_CORRELATION/GLCM_INFOMEAS1 are the
  soft-NAN(=0)-guarded ones on the single degenerate slice.
• The map/function were renamed (glcm_values, glcm_truth_key) as part of this GLCM test's own
  refresh; this is GLCM-local, not the cross-cutting table-merge refactor (which was dropped).

External-oracle proof lives in the Python tests

The phantom goldens are regression anchors; the oracle validation is in
tests/python/test_feature_bugs.py, which runs the production featurize() path:

• test_glcm_contrast_nonzero_by_default — a horizontal intensity ramp must have 0° contrast > 0 and
  90° contrast == 0. Directly catches bug (1) (offset=0 → all-zero contrast).
• test_glcm_background_not_counted — on the canonical concave ROI quantized to levels 1..64 (identity
  binning), GLCM_CONTRAST_AVE must match the transparent numpy reference / MIRP / PyRadiomics
  (≈133.9). Pre-fix this ROI gave ≈99 (background-diluted). Catches bug (2).

test_glcm_bugs.h (new) + test_all.cc

test_glcm_bug_offset_default_is_one() is a C++ regression guard that builds GLCM through the
production settings path (not the hard-coded phantom), asserting the co-occurrence distance
defaults to 1 — the exact gap that let bug (1) ship. Registered in test_all.cc as
TEST_GLCM_BUG_OFFSET_DEFAULT (only the #include and the TEST(...) block were added; the file's
pre-existing mixed line endings were left byte-for-byte untouched — no churn).

CI

Full suite green on this branch: C++ runAllTests 683/683 passed; pytest tests/python/
50 passed, 1 skipped (48 baseline + the 2 GLCM tests above).