nx_superopt_race_test.nx

nx_superopt_race_test.nx -- the team SYNTHESIZES + SUPEROPTIMIZES + EMITS real machine code, proven 1:1 by EXECUTION, ready to race gcc -O2. For each kernel x*C: 1. so_find synthesizes the SHORTEST program (provably minimal over its ISA) 2. se_emit_full emits competitive x86-64 (lea fusion + in-place reuse) 3. eng_link + eng_run assemble and RUN it; the exit byte is sum(synth(d)) over HELD-OUT inputs d -- compared to (sum d*C)&255, so a match proves BOTH that the synthesis generalizes (not overfit) AND that the emit is faithful. Prints the team's instruction count per kernel (the number for the gcc -O2 race). Gate: every kernel verifies 1:1 -> exit 0.