docs: 3-A gate result — KHOA model tide viable on large-tide coasts

Ran the advisor's verification gate (12 stations, today, official high/low
vs model zeta extremum timing). Findings:
- pre_date is an integer day-index (0..6), not a date string — page JS
  confirms `pre_date: index`; index 3 = today, up to +3 days forward.
  This resolves the earlier "stale 6/2" concern (date-lock was solved).
- Model tide timing matches official within 12-30 min median on West/South/
  Jeju (range >= ~1.5m): 인천/군산/목포/여수/통영/서귀포 all PASS.
- East-coast small-tide (묵호 24cm/후포 20cm/포항 17cm) FAILS: 94-138 min
  error, 12-17 spurious extrema/day — model noise floor (~±0.3m) buries the
  ~0.2m signal. Needs an automatic guard (daily-range/extrema-count) + fall
  back to nearest official station.
- Bonus: 7-day endpoint serves current/temp/salinity forward, enabling the
  backlog 조류세기 + 수온 features from the same source.

Constraint: KHOA ocean-grid is unofficial; arbitrary-coordinate tide is coast-conditional (large-tide only) and must degrade to the 61 official stations.
Confidence: high
Scope-risk: local
Reversibility: trivial
Directive: If we build this, implement the noise guard first; keep official 61 stations as the absolute-cm source of truth.
Tested: Live gate against data.go.kr official high/low oracle, 12 stations, 2026-06-05.
Not-tested: Bay-interior non-official points (no oracle); multi-day week reproduction; guard threshold tuning.
Related: docs/BACKLOG.md

🐙 Autopus <noreply@autopus.co>