fix(chords): correct max-chord angle index and all-chords histogram s…

[darkclad]

…ource

ChordsFeature::calculate computed the max-chord angle index from iteMin (idxmax = distance(begin, iteMin)), so MAXCHORDS_MAX_ANG always equalled MAXCHORDS_MIN_ANG (same copy-paste for ALLCHORDS). It also built the all-chords mode/median histogram from the max-chords vector (MC) instead of all chords (AC).

Index the max-angle from iteMax and initialise the all-chords histogram from AC. Refresh the vetted chord goldens in test_2d_remaining_features.h (MAXCHORDS_MAX_ANG and ALLCHORDS_MAX_ANG -> 0.0; ALLCHORDS_MEDIAN/MODE -> 3.0) and add tests/python/test_feature_bugs.py::test_chord_max_angle_distinct_from_min.

[darkclad]

Justification — golden value changes in tests/test_2d_remaining_features.h

Branch: main-chords-max-angle (split from main-feature-validation, based on main)
Source fix: src/nyx/features/chords.cpp
Test file touched: tests/test_2d_remaining_features.h (4 golden values)

The two bugs in chords.cpp

1. idxmax used iteMin — the max-angle index was taken from the minimum chord
   (idxmax = std::distance(begin, iteMin)), so the reported "max-angle" was actually the
   min chord's angle. Result: MAXCHORDS_MAX_ANG was numerically identical to
   MAXCHORDS_MIN_ANG (same copy-paste error for the ALLCHORDS block).
2. histo.initialize_uniques(MC) — the all-chords mode/median histogram was built from the
   max-chords vector MC instead of the all-chords vector AC. Result: ALLCHORDS_MODE
   and ALLCHORDS_MEDIAN reported the max-chords mode/median.

Justification per changed value

Golden	Old	New	Why the new value is correct
MAXCHORDS_MAX_ANG	0.94247779607693793	0.0	Old value ==MAXCHORDS_MIN_ANG (0.9425) — the bug signature (max-angle = min-angle). After the fix it is the angle of the longest max-chord, which is horizontal → 0.0 rad. Consistent with MAXCHORDS_MAX = 6.0.
ALLCHORDS_MAX_ANG	0.15707963267948966	0.0	Same signature: old ==ALLCHORDS_MIN_ANG (0.1571). Longest all-chord is horizontal → 0.0 rad.
ALLCHORDS_MEDIAN	4.0	3.0	4.0 was themax-chords median (leaked from MC). The true all-chords median must be consistent with ALLCHORDS_MEAN = 2.9134615…, which was always computed correctly from AC and did not change. Median 3.0 ≈ mean 2.913 for a right-skewed short-chord distribution; median 4.0 > mean 2.913 was internally inconsistent.
ALLCHORDS_MODE	4.0	3.0	SameMC leak; the AC length distribution (min 1, max 6, mean 2.913) peaks at 3.0.

Why this is not fudging

• What did NOT change is the tell. These goldens are unchanged because their code paths were
  already correct: MAXCHORDS_MAX (6.0), MAXCHORDS_MIN (3.0), MAXCHORDS_MEDIAN (4.0),
  MAXCHORDS_MEAN, MAXCHORDS_MODE, MAXCHORDS_STDDEV, ALLCHORDS_MAX (6.0), ALLCHORDS_MIN (1.0),
  ALLCHORDS_MEAN (2.9134615…), ALLCHORDS_STDDEV, MAXCHORDS_MIN_ANG (0.9425),
  ALLCHORDS_MIN_ANG (0.1571). Only the 4 outputs of the two buggy lines moved — a fudge would ripple.
• The values are the fixed code's deterministic output, not hand-picked. The C++ gtest
  TEST_NYXUS.TEST_REMAINING2D_VERIFIABLE_WITH_3P_BUILTIN_ORACLE_CHORD_STAT_FEATURES and
  ..._UNVETTED_NO_DIRECT_ORACLE_CHORD_ANGLE_FEATURES pass (tolerance 1e-3) with exactly these
  numbers after rebuilding — i.e. the corrected binary computes 0.0 / 0.0 / 3.0 / 3.0.
• They are byte-identical to the oracle-validated main-feature-validation values — copied, not
  invented. (Only the two MAX_ANG and two ALLCHORDS MEDIAN/MODE numbers were taken; the shared
  table-merge refactor and unrelated radial-distribution vector goldens in that file were intentionally
  left out — those belong to the contour-frame cluster branch.)

Regression guard

tests/python/test_feature_bugs.py::test_chord_max_angle_distinct_from_min asserts, on the canonical
ROI, that MAXCHORDS_MAX_ANG != MAXCHORDS_MIN_ANG and ALLCHORDS_MAX_ANG != ALLCHORDS_MIN_ANG
whenever the max and min chord lengths differ — directly catching bug (1) if it regresses.

CI

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