test(moments): assert the normalized raw moments against scikit-image

[darkclad]

32 rows of the 2D moments family - NORM_SPAT_MOMENT_ij and IMOM_NRM_ij - carried status=vetted with source=tracker while no in-tree test asserted them against any oracle. They now have one.

Both are normalized RAW moments, m_pq / m_00^((p+q)/2 + 1). skimage has no native function for that: moments_normalized() applies the same exponent to the CENTRAL moments, which is a different quantity and one Nyxus already exposes separately as NORM_CENTRAL_MOMENT_pq / IMOM_NCM_pq. On the pinned 48x40 fixture the two differ plainly - 0.3876 against 0.0999 at p=2,q=0 - and the second is what NORM_CENTRAL_MOMENT_20 matches.

So the oracle is skimage's raw moments carried through skimage's own normalization exponent: the moment sums, which are the non-trivial part, come from an independent implementation, and the only step layered on top is an exponent skimage itself uses. Nyxus reproduces all 32 exactly - worst relative difference 0.0e+00 - and the generator now asserts that normalized-raw and moments_normalized are numerically distinct, so a later edit cannot silently swap one for the other.

• gen_moments_skimage.py gains section D, emitting the 32 goldens plus that guard.
• test_2d_moments_skimage.h gains two ref-val tables and test_2d_moments_normraw_{shape,intensity}_skimage, registered in test_all.cc.
• oracle_coverage.csv: all 180 rows move to source=in-tree with config_recipe=moments.skimage_regionprops and tolerance=rel=1e-6, and current_test is now derived from what the tests actually assert rather than from what the tracker-era registry claimed. That correction alone touched 72 rows in both directions: rows naming the regression file that it does not cover, and the 32 that it does cover but which did not say so.

The 62 weighted (distance-to-contour) rows stay regression, and now carry the reason in-tree rather than only in a side document. Re-measured against current code: the weighting takes dist from Pixel2::min_sqdist, an approximate hill-descent its own doc comment warns overestimates, and on the pinned fixture the exact distance to the contour gives WEIGHTED_SPAT_MOMENT_00 = 1821.28398639964 against the pinned 2498.7137504625134 - 1.372x, a 37.2% overestimate. Fixing that changes 62 published feature values and swaps an O(pixels x contour) scan for a distance transform, so it is out of scope here and noted for its own branch. One correction for whoever takes it: there is exactly one live weighting path; the second implementation using a different formula (inten/(dist+eps), image_matrix.cpp) has no callers anywhere, and no GPU kernel does this weighting at all.

Also, per tests/vetting/ housekeeping:

• Include hygiene: test_2d_moments_common.h called init_label_record_3 / update_label_record_3 unqualified, so all three moments headers compiled only because another header earlier in test_all.cc's include order contributes a file-scope using namespace Nyxus;. All three now compile as the only thing in a translation unit. Its unused is dropped.
• The generator re-verified a hand-picked list of 25 golden names against the fresh run. It now re-verifies EVERY golden pinned in test_2d_moments_skimage.h - 142 of 142, across both fixtures, none without a skimage counterpart - so the check cannot quietly stop covering entries added later. Doing so surfaced that moments_central(order=N) leaves entries with p+q > N at zero while moments() fills the whole matrix, so the verification computes its matrices at order 6; the tables pin up to CENTRAL_MOMENT_33.
• Comments describing the history of an h5/h6 formula fix are replaced by statements of the current contract - what the value is, and what a wrong formula would produce instead - with the history left to the audit report.
• New audit artifacts: moments_2d_coverage.csv, moments_2d_skimage_vetting_report.md, moments_2d_golden_regen.md, and scan_moments_coverage.py, whose --check regenerates the CSV from the test sources and enforces that the registry and the tests agree row for row with no unregistered gtest function.
• TOOLS.md records the moments_normalized-vs-normalized-raw trap and the p-on-x coordinate convention.

gtest 734/734; ASan+UBSan on Linux 730/730 with no sanitizer diagnostics; pytest 85 passed, 1 skipped, 7 arrow-only deselected; check_test_names, check_coverage and scan_moments_coverage all clean.