450 runs/arm x 2 arms (15 opponents x 30 runs x 3 rounds, conc=6, 0 failed,
0 never started, 0 env mis-set), frozen binary d9a39c3b8472, TR_MOVEMENT=tfil.
Round-win 40.30% -> 39.70% (-0.59 pp, CI -3.90..+2.72, sign-flip exact-2^15
p=0.7676, MDE 4.73 pp); wins/run 0.200 -> 0.182 (MDE 0.1420); damage/run
113.65 -> 112.70 (p=0.5298, MDE 4.19). Deviation disclosed: 30 runs/opponent
instead of 42 (throughput 14-22 runs/min vs 22.7-23.2 assumed); full 15-opponent
panel and both arms kept, runs reduced.
Headline is the mechanism failure: firing was suppressed on 0.04% of ticks,
not the 9.6% the offline ruler predicted (~200x smaller); only 4.8% of shots
happen in the low-energy zone and the floor removed 14% of those; rounds ending
at self energy <=0 were 40.6% live vs 61.2% implied by the offline corpus;
median self energy at death 14.1 -> 15.5. The pre-registered 'we cannot measure
the damage cost directly' call was correct.
DO NOT ADOPT. TR_RAM_FLOOR_ENERGY stays 0.0, do not re-test. The null does not
prove the knob inert - the mechanism barely fired.
SirStone
51bfa57067