w2heat_cs14 — the champion: broadly beats the King, wins every seed

Our strongest generator. Same engine as worker_1/w2heat, with conditional_stability at its tuned peak of 0.14 (config-only). It is the top of a clean dose-response and beats the reigning King broadly and consistently at mainnet scale.

What it is

Rich cascade-v19 16-family engine + demand @0.10 + the heat grafts (gapless coverage, count/discrete tail tilt, and conditional_stability — alternating long calm / short dynamic segments teaching "stay flat on calm targets"). That lever both wins the geometric-MASE heat AND cuts per-round duel variance; the sweet spot on our base is 0.14.

Validation — FIXED Toto2-4m, 8-seed 2B (mainnet scale) vs jan/cascade9 (King)

conditional_stability dose-response (mean LCB): 0.07 → −0.049 · 0.11 → +0.014 · 0.14 → +0.060 (peak) · 0.18 → −0.013 (overshoot).

At the 0.14 peak:

• P_win 8/8 — beats cascade9's point estimate on every seed.
• worst-case LCB −0.018 — never catastrophically collapses (vs the base w2heat's −0.441 and cs=0.11's −0.260).
• mean LCB +0.060, dethrones (LCB > +0.02) on 5/8 rounds.
• The base already wins BOTH score halves at 2B (MWSQL 0.968 / MASE 0.901, composite ~+6.6% vs cascade9, 6/7 domains); cs=0.14 adds the consistency.
• Throughput ~0.79× King; cascade verify OK, deterministic (corpus_digest de7d17a1…); content-dedup safe (large token delta from cascade9 / jtest-v26 — no exact-identity collision).

Honest scope

This genuinely revises the earlier "~1/4-3/8 dethrone lottery is a hard cap" claim: at the tuned conditional_stability dose the model beats the King on every seed and clears the dethrone bar most rounds, with no catastrophic downside. It is still not a guaranteed 8/8 dethrone — 3 rounds land in the +0-to-+0.02 band — so the play remains field-and-persist, but from a position of broad + consistent superiority over cascade9, not a coin flip.

Submit as generator.Generator.