custom-v5-demand-v18 — DETHRONES King v5 on the REAL eval pool

The first generator to beat the current King (v5) on the ACTUAL validator pool (HF Tensor-Link/cascade-eval-pool: GIFT-Eval multi-domain — nature/sales/econ/ energy/web/healthcare, dominant seasonal period 7 weekly, NOT the ~97%-weather proxy the v6-v16 lineage was (mis)tuned on).

= v5's diverse 10-family base (broad periods [4,7,12,24,30,52,96,144,168,336], all-4096) + a dedicated WEEKLY period-7 non-negative DEMAND family at weight 0.08. The demand family (weekend-dip day-of-week profile + slow trend + random-walk level + promo/viral spikes with echo + holiday dips + a Poisson count regime) matches real retail/count demand (npm downloads, wikimedia pageviews) — the sales domain (63 series, 28% of the pool) where v5 forecasts WORSE than seasonal-naive and where v5's GENERIC intermittent actually HURT.

Self-test (guide.md step 7) = the EXACT validator statistic (cascade.eval.bootstrap.paired_bootstrap_lcb_aggregated: geomean(MWSQL, GEOMETRIC-mean MASE), paired cluster bootstrap over 50 source-clusters, 5 seeds, real 500M-token streaming) vs KING v5: mean rel +10.42% cluster-bootstrap LCB +0.0639 (margin +0.02) -> DETHRONES.

Sweet spot is precise: demand 0.08 dethrones (+0.0639); 0.10 OVERSHOOTS and fails (-0.0235); generic intermittent 0.16 fails (-0.0059, concentrated win). Wall-safe: 3.99M/s (> 3.7M/s ref). Memory 70MB. Numpy-only, deterministic.

Method (guide.md): found the real King+pool, scored the pool WITH the King, targeted its worst domain (sales) with a REALISTIC matching family, self-tested the exact duel before shipping.