w2heat_cs10 — heat-competitive + King-beating strike model

Identical generator code to worker_1/w2heat, with conditional_stability dosed up 0.07 → 0.11 (config-only). That single change is a large, monotonic win at mainnet scale — it is our strongest generator to date.

What it is

The rich cascade-v19 16-family engine + our demand family @0.10, plus the heat grafts (gapless coverage, count/discrete tail tilt, and conditional_ stability — alternating long calm / short dynamic segments that teach "stay flat on calm targets," removing the flat-target worst-case MASE). At conditional_stability = 0.11 that lever also cuts per-round duel variance.

Validation (FIXED Toto2-4m, real held-out pool, vs jan/cascade9 = current King)

• 8-seed 2B (mainnet-scale FULL LCB): dethrones 6/8, mean LCB +0.014 (positive), worst −0.260, P_win 6/8 — vs w2heat's cs=0.07 (3/8, mean −0.049). Doubling the dethrone rate and crossing the mean positive.
• At 2B the base w2heat already wins BOTH score halves vs cascade9: MWSQL 0.968, MASE 0.901, composite 0.934 = ~6.6% better overall, winning 6 of 7 domains — cs=0.11 lifts the per-round consistency on top of that.
• Throughput: ~0.79× King (9.2 M pts/s) — well above v30's fatal 0.64×.
• Deploy checks: cascade verify OK, deterministic; content-dedup safe (generator code carries a large token delta from cascade9 / jtest-v26 / the rich base — no exact-identity collision).

Honest scope

conditional_stability dose genuinely lifts the single-round dethrone rate past the ~3/8 treasury-lottery band I previously believed was a hard cap. A worst-case seed (−0.26) remains, so it is not a clean 8/8. The optimal dose is being swept (0.11 confirmed strong; 0.14 / 0.18 under test) — if a higher dose keeps climbing without overshoot, prefer that variant. Winning strategy is still consistency + persistence, but this model wins most rounds at scale, not ~1/4.

Submit as generator.Generator.