feat(proxy): near-station snap, night moonrise, relaxed prominence

Evidence-based response to the tide-accuracy ask. Measured SIGNED bias
(model − official) at 8 large-tide stations on 6/6: scattered from −68
(완도, early) to +23 (인천, late), median ~±5min, with per-station sign
flips. So a time-shift / 조시차 correction is NOT valid (it would overfit
the reference app, which isn't ground truth) — not built. Instead:
- routes/tide: coordinate spots within 6km of an official station now serve
  that station's exact measured tide (same water, sub-hour precision) — e.g.
  통영 coord recovers the 00:06 high the hourly model had dropped.
- tideExtrema: relaxed prominence (0.12→0.09*range, floor 0.12m) to stop
  over-pruning real secondary highs/lows.
- astro: when moonrise/moonset fall past local midnight (null), fall back to
  the next day's first occurrence so night moonrise is shown (user-requested;
  fishing needs it).

Verified: 통영 coord -> official exact; 노력항 model -> clean 4 events;
moonrise no longer null. vitest 28/28.

Constraint: no synthetic time correction — model timing is left at its honest ~±5–30min vs the 조시차 standard (hourly-sampling floor is irreducible via this endpoint, which rejects sub-hour interval params).
Confidence: high
Scope-risk: module
Reversibility: trivial
Directive: To truly match 물때표 on named 부속항 we'd need KHOA 조시차 tables (non-public); the near-station snap covers ports close to the 61 stations exactly.
Tested: vitest 28/28; live HTTP — 통영 snap, 노력항 model, moonrise fallback; signed-bias measurement (scattered) recorded.
Not-tested: snap distance threshold tuning (6km chosen conservatively).
Related: docs/BACKLOG.md 3-A

🐙 Autopus <noreply@autopus.co>