Classify hybrid (annular-total) eclipses; add NASA 2031 Nov 14 hybrid reference (item 14)

[wiyth00]

Summary

Work-plan item 14. enumerate_real_solar_eclipses labelled every central eclipse by the sign of the near-side core radius at greatest eclipse only, so genuinely hybrid eclipses were mislabelled total (or annular). The new _classify_central_eclipse helper samples the signed core radius along the whole central track (±3 h at 120 s steps via iter_real_track, off-ellipsoid samples skipped); if the radius takes both signs anywhere on the track the eclipse is labelled hybrid, matching NASA/GSFC catalog conventions. Otherwise the label falls back to the previous sign-at-greatest rule, unchanged.

Adds the NASA/GSFC 2031 Nov 14 hybrid eclipse (Saros 143, member 24/72) as a validation reference, appended at the end of NASA_GSFC_REFERENCES to keep the tuple chronological. All values are quoted verbatim from the GSFC Besselian-elements page: greatest eclipse 21:07:30.2 TDT (= 21:06:12.3 UT), ΔT 77.9 s, 0°38.0′S 137°37.8′W, central duration 01m08.3s. The existing parametrized tests in tests/test_known_eclipses.py automatically pick this reference up, so DE440s timing (<60 s TT), greatest-eclipse position (<25 km), and the hybrid type label are all CI-enforced.

New tests/test_hybrid_classification.py:

• test_2031_nov_14_is_classified_hybrid — fails on main (main labels it total).
• Guards that 2026-08-12 stays total and 2027-02-06 stays annular (no relabelling of unambiguous events).

Scientific impact

• No change to computed results
• Changes computed results (explain below)

Eclipse type labels move for genuinely hybrid rows in the real-Moon catalog (timing, coordinates, and geometry are untouched — the classifier only reads additional samples). Two caveats for review:

1. Please re-run the full catalog and diff the eclipse_type column of real_moon_eclipses.csv against a pre-change run to see exactly which rows move; any downstream logic that filters on total/annular labels would see those rows reclassified.
2. The 2026-08-12 anchor event is unaffected (regression-tested).

Runtime: each central eclipse now costs up to ~180 extra shadow evaluations (one per 120 s track sample, short-circuiting once both signs are seen). Adding the 2031 reference also widens the validation enumeration window from 2026–2028 to 2026–2031, so the validation stage and tests/test_known_eclipses.py take proportionally longer.

Checklist

• Regression test fails on main (test_2031_nov_14_is_classified_hybrid).
• Reference values quoted verbatim from the GSFC source page (no fabricated digits).
• CI green — note: CI is the first integrated execution; I could not run pytest locally in this environment. Will fix promptly if red.

Note for merge ordering: this PR appends to the end of NASA_GSFC_REFERENCES; the item-17 PR inserts at the beginning and both keep the tuple chronological, so the two merge cleanly in either order.