custom-wbp-cloud-v16fast — submission candidate (v16 quality, wall-safe generation)

v16's EXACT data-prior (radiation night-floor + cloud archetype + weather family + 10-family mix, all-4096, +14.02% vs the v2 baseline) but generation-optimized to clear the mainnet 3.7M tok/s wall:

• per-archetype SLICED dispatch (compute cloud only on its ~18% slice, not all series)
• block-scan AR(1) (Blelloch segmented scan; ~160 iters vs 4096; rel-err 3e-16 vs the loop) Measured 4.36M tok/s (v16 = 2.83M/s, below the wall). Distribution verified == v16 within re-seed noise (cloud-fraction, per-series moments). Net vs v16 on mainnet: ~+30% MORE trained tokens in the 3h wall (v16 starves to ~78% of the 40B budget), same priors => strictly >= v16. Numpy-only, deterministic, _sanitize-gated.