feat(proxy): tide quality — prominence filter, midnight stitch, level numbers

Addresses the accuracy issues found comparing 장환/고마리 against the
reference app:
- tideExtrema: cleanExtrema() prunes hourly-sampling wiggles by prominence
  (< max(0.15m, 0.12*range)) and collapses same-type runs — removes the
  spurious midday high/low (e.g. 6/5 14:00). Noise guard still uses raw
  extrema count so East-coast small-tide still degrades.
- khoaModel: fetchModelTideDay now stitches the previous/next day's zeta
  (±1 day) so near-midnight highs/lows are detected (6/5 00:40 high now
  caught at 00:22). Noise judged on the target day only so the stitch
  doesn't inflate the count.
- levels.ts annotateLevels(): per-event delta (▲/▼ vs previous extremum)
  and daily range — datum-independent, so reliable even for the MSL-relative
  model. Applied to both official and model tides; TideEvent.delta +
  TideDay.rangeCm added.

Verified: 장환 6/5 -> clean 4 events (만조 00:22, 간조 06:55 ▼247, 만조 12:15
▲162, 간조 18:41 ▼233), range 318cm; 6/6 clean. vitest 28/28.

Constraint: model levels stay MSL-relative; only deltas/range are shown as absolute cm because differences are datum-independent.
Confidence: high
Scope-risk: module
Reversibility: moderate
Directive: Remaining ~15min early bias on bay points is a separate task (nearest-station time-offset correction); not addressed here.
Tested: vitest 28/28; live HTTP 장환 6/5·6/6 (spurious extremum gone, midnight high recovered, deltas/range present).
Not-tested: prominence threshold on diurnal-inequality coasts (could over-prune a real secondary); revisit if reported.
Related: docs/BACKLOG.md 3-A/5

🐙 Autopus <noreply@autopus.co>