Track 1: directions that did not beat the best

The bar: F3t 0xE4B713F9 (E2b + G5 attention overlap + timing hardening): Llama 831.14 tok/s @ 5.89 [M] (+2.0% vs same-day E2b 812.22-816.61), Qwen 249.62 @ 7.14 [M] (within noise); G5 751.6 to 520.0 us per tree, tree -8.3% [M]. Every row below was measured and did not beat it. 2026-11-09, box1-qwen-junior.

Calculations and hypothesis: when each could shine again

#DirectionImage / testMeasured [M]Why it failed

Source: reports/results/DRAFTER_SPEED_RESEARCH_20261107/FAILED_DIRECTIONS_20261109.md and the docs it cites; record evidence in reports/results/TRACK1_RECORD_F3T_20261109/.