custom-wbp-cloud-v16 — CURRENT CHAMPION (supersedes v13)

v13 (radiation night-floor) + a disjoint ~18% bounded synoptic-saturating cloud_cover archetype. Generalizes v13's "match the real feed's structure" principle to the 2nd-highest-error feed. Real cloud_cover is bounded [0,100] with ~52% of hours pinned at a boundary (clear/overcast), synoptic red-noise driven, rough/heavy-tailed increments, only weak diurnal — v13's generic smooth-diurnal weather matched none of it. The cloud archetype (multi-timescale red noise hard-clipped to [0,100], per-series variety) teaches bounded support + saturation persistence.

Real 500M budget, production-faithful streaming, honest 60-feed clustering:

• vs v2 king: +14.02%, LCB +0.0999, 8/8 seeds
• vs v13 DIRECT: +1.34%, 8/8 seeds positive (LCB +0.0082 — genuine but below the 0.02 dethrone margin: a real small top-up, not a landslide)
• per-variable: cloud_cover +0.8% (v13) -> +5.3% (v16); radiation floor intact (+27.2%)

Exhausted around here: wind and humidity archetypes WASH (no net gain on top of cloud; see v17/v18). Radiation remains the one big deterministic lever; cloud is the last genuine per-feed top-up. Numpy-only, deterministic, all-4096. Honest caveat inherited from v13: radiation margin is a one-shot dethroning margin vs a radiation-naive king.