fix(convex-hull): Pick's-theorem hull area so SOLIDITY <= 1

[darkclad]

ConvexHullFeature computed the hull area with the bare shoelace polygon area, which runs through pixel centres and under-counts the rasterised pixel coverage. The ROI area is a pixel count, so SOLIDITY = roi_area / hull_area could exceed 1 (1.3 on the canonical ROI) - impossible.

Use the Pick's-theorem pixel-count-equivalent hull area (shoelace + boundary_points/2 + 1, boundary_points = sum of gcd(|dx|,|dy|) over edges), matching scikit-image convex_area, plus a hull>0 guard. SOLIDITY is now <= 1 in both the trivial and out-of-core paths.

Refresh the vetted goldens in test_shape_morphology_2d.h (CONVEX_HULL_AREA 20->27, SOLIDITY 1.3->0.96296 = 26/27; CIRCULARITY and all other shape goldens unchanged) and add tests/python/test_feature_bugs.py::test_solidity_le_one.

[darkclad]

Justification — golden value changes for main-convex-hull-solidity

Branch: main-convex-hull-solidity (split from main-feature-validation, based on main)
Source fix: src/nyx/features/convex_hull_nontriv.cpp
Test file touched: tests/test_shape_morphology_2d.h (2 goldens: CONVEX_HULL_AREA, SOLIDITY)

The bug

ConvexHullFeature::calculate computed the hull area with the bare shoelace polygon area
(polygon_area(convHull)), which runs through pixel centres and therefore under-counts the
rasterised pixel coverage of the hull. The ROI area (s_roi) is a pixel count. Dividing a
pixel-count ROI area by a centre-based hull area gives SOLIDITY = s_roi / s_hull > 1 for
small/elongated ROIs — impossible (solidity is by definition ≤ 1). On the canonical fixture the old
golden was SOLIDITY = 1.3.

The fix

Use the Pick's-theorem pixel-count-equivalent hull area so the hull is measured on the same basis
as the ROI (pixel count):

s_hull = shoelace_area + boundary_points/2 + 1      // boundary_points = sum over edges of gcd(|dx|,|dy|)

This matches scikit-image's convex_area (the pixel count of the rasterised hull). Result:
SOLIDITY = s_roi / s_hull ≤ 1, plus a s_hull > 0 guard.

Golden changes (2 keys)

Golden	Old	New	Why
CONVEX_HULL_AREA	20.0	27.0	Pick's-theorem pixel-count hull area on the fixture. External oracle scikit-imageconvex_area = 28 (≈4% high; the two hulls differ by one boundary lattice point).
SOLIDITY	1.3	0.9629629629629629	= AREA_PIXELS_COUNT / CONVEX_HULL_AREA = 26 / 27 = 0.962962…. scikit-image gives 26/28 = 0.929. Now ≤ 1.

Why this is not fudging

• Internal consistency: the new SOLIDITY is exactly AREA_PIXELS_COUNT (26) / CONVEX_HULL_AREA (27); both the numerator (26, unchanged golden) and the new denominator are in the same table.
• Nothing else moved: the fix only changes s_hull. CIRCULARITY (which uses s_roi and the
  perimeter, not the hull) is unchanged, as are all other shape goldens. Only the 2 hull-derived
  keys changed.
• Sign of correctness: the old value 1.3 was provably wrong (solidity can never exceed 1); the new
  value is ≤ 1 and within ~4% of scikit-image's independent convex_area/solidity.
• Values are byte-identical to the oracle-validated main-feature-validation (copied, not invented);
  only the shared table-merge refactor was dropped (I kept main's two-table structure and edited the
  two numbers in place).

Regression guard

tests/python/test_feature_bugs.py::test_solidity_le_one asserts SOLIDITY <= 1 + 1e-6 on the
canonical ROI through the production featurize() path — the hard correctness invariant that the old
code violated (1.3).

CI

Full suite green on this branch: C++ runAllTests 683/683; pytest tests/python/ 49 passed, 1 skipped.