nx_race_vs_c_test.nx

nx_race_vs_c_test.nx -- the ONGOING racing gate: Nishi vs C, 1:1, sovereign. "win over C in an ONGOING fashion" (operator) -- so this is a PERMANENT gate, not a one-shot. It forks+execs both pre-built contestants (the SAME kernel compiled by our nx_cc and by gcc -O2), captures each one's "<checksum> <cycles>" line, and judges. The HARD invariant is CAPABILITY: byte-identical checksum = our codegen did NOT miscompile. Speed/size are REPORTED + tracked, never red-failed -- we already KNOW we trail gcc on raw compute (the stack-machine/no-regalloc tax), and a gate that fails on the known grind would be a false-failing gate (racing doctrine: bake a NON-false-failing gate). It fails ONLY on a correctness break -- the one thing that must never regress while we close the speed gap. Build lane must produce /tmp/race_nishi.elf and /tmp/race_c.elf first (the two contestants). Known answer: byte-identical checksum -> exit 0; prints the board. CITATIONS (real): fair-benchmark methodology -- Curtsinger & Berger, "Stabilizer: Statistically Sound Performance Evaluation," ASPLOS 2013. The energy/time/memory tradeoff this race feeds into -- Pereira et al., "Energy Efficiency across Programming Languages," ACM SLE 2017. The named speed lever -- Poletto & Sarkar, linear-scan register allocation, ACM TOPLAS 1999.