perf: build the real-Moon centerline once per real eclipse in fixed_system_search

[wiyth00]

Summary

In fixed_system_search, generate_central_track(context, real_tt, "real_moon", step_seconds=120.0, include_partial_maximum=True) sat inside the for second_event in second_events: loop. Its only time-dependent argument is real_tt, which is fixed for the whole loop, so the identical track was recomputed for every second-moon candidate of the same real eclipse.

• Hoist the real-Moon generate_central_track call above the inner loop
• Add if not second_events: continue first, so real eclipses with no candidates do less work than before rather than more (without this guard, hoisting would build a track for every real eclipse in the window, including the ones that currently build none)

Work plan item P1 / 4.

Note on aliasing

Candidates from the same real eclipse now share one real_track dict rather than each holding an equal-valued copy, including the entries stored in tracks[event.event_id]. Every downstream consumer (minimum_track_distance_km, _best_common_location, plotting) reads without mutating, so this is safe — but it is the one behavioural difference worth a reviewer's eye.

Scientific impact

None. The hoisted call is argument-for-argument identical to the one it replaces, and generate_central_track is a pure function of (context, tt_jd, body, step_seconds, include_partial_maximum). Track distances, classifications, and site selection are bit-for-bit unchanged.

Suggested verification per the work plan: run the fixed-system search over a 1-year window and diff all_candidates.csv before/after — it should be byte-identical.

• No change to computed results
• Results change (described above)

Checklist

• ruff check . passes
• pytest passes locally
• New or changed behaviour is covered by a test
• Docs / README updated if user-facing behaviour changed

tests/test_fixed_system_search_track_reuse.py stubs the search's collaborators so it needs no ephemeris kernel, and asserts one real_moon build against three second_moon builds. It fails on main (3 real-Moon builds) and passes here. A second case asserts no track is built when a real eclipse yields no candidates.

Lint/test boxes left unchecked because this branch was pushed without a local run — CI on this PR is the first execution.