Expand validation references: 2024 pre-epoch total + 2025 timing-only partial (item 17)

[wiyth00]

Summary

Work-plan item 17. Extends NASA_GSFC_REFERENCES beyond the model epoch and beyond central eclipses, with all values quoted verbatim from GSFC Besselian-elements pages:

• 2024 Apr 08 total (Saros 139) — pre-epoch historical anchor, inserted at the beginning of the tuple (chronological order). Greatest eclipse 18:18:29.0 TDT (= 18:17:18.3 UT), 25°17.2′N 104°08.3′W, duration 04m28.1s. Note: the page states ΔT = 70.6 s while its own TDT/UT pair differs by 70.7 s — one 0.1 s rounding step. Both values are kept verbatim; the fixture self-consistency test tolerance is relaxed from 0.051 s to 0.105 s with a comment explaining exactly why.
• 2025 Mar 29 deep partial (Saros 149, gamma 1.0404, magnitude 0.9378) — non-central. NASA publishes no greatest-eclipse coordinates or central duration for non-central eclipses, so the reference stores None for those fields and is validated on eclipse type + TT timing only. This scope limitation is explicit rather than papered over: the dataclass docstring, a new non_central_reference_note in the validation report, and tests/test_known_eclipses.py skips all say so.

Mechanics:

• PublishedEclipseReference.latitude_deg/longitude_deg/central_duration_s become float | None; EclipseValidationResult coordinate/duration fields become None-aware and coordinate_within_target becomes bool | None. _validate_one skips the WGS84 distance and local-circumstances solve for timing-only references; passed = type_matches and timing_within_target and coordinate_within_target is not False. Central references matched to a non-central event still raise ValueError (regression-tested).
• One small cli.py guard: the run_full terminal summary's “max path error” line now computes over central references only, so a coordinate-less reference can't crash or poison the summary. Everything else in cli.py is byte-identical to main.
• New tests/test_validation_references.py (partial reference is genuinely timing-only; _validate_one behavior on a fabricated non-central event — fails on main, where it raises ValueError; central references still require central events) and test_known_eclipses.py updates (coordinate test skips non-central refs; new chronological-order test that enforces the insert-at-front/append-at-end discipline shared with the item-14 PR Classify hybrid (annular-total) eclipses; add NASA 2031 Nov 14 hybrid reference (item 14) #15).

Scientific impact

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

No physics or geometry changes. But the validation enumeration window grows to 2024–2028 (2024–2031 once PR #15 also lands), so validation_report.json / validation_results.csv gain rows and the validation stage + reference tests take proportionally longer in CI and in full runs. Partial-reference rows carry null coordinate/duration fields — downstream consumers of the report should tolerate that (the shipped CLI summary already does).

Checklist

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

Merge-order independence vs. PR #15 (item 14): this PR inserts at the beginning of the reference tuple and touches _validate_one/result dataclasses; #15 appends at the end and touches the classifier in ephemeris.py. The single cli.py hunk here is also disjoint from PR #14's logging changes. All three merge cleanly in any order.