nx_regalloc_race_kernel_test.nx

nx_regalloc_race_kernel_test.nx -- point the proven allocator at the EXACT hot loop we lose ~1.9x to C on (nx_race_kernel), to quantify the lever concretely. The race kernel's loop body, as a straight-line IR (each line defines one vreg): c_in, acc_in, i_in live-ins (already in registers) t0 = c_in * K (3) c1 = t0 + A (4) t1 = c1 >> 31 (5) c2 = c1 ^ t1 (6) <- next iteration's c (live-out) t2 = c2 & 65535 (7) acc1 = acc_in + t2 (8) <- next iteration's acc (live-out) i1 = i_in + 1 (9) <- next iteration's i (live-out) cmp = i1 < K (10) Pressure stays low (~4-5 live), so on x86-64's ~8+ usable GP registers the WHOLE body fits in registers: ZERO memory traffic. The stack-resident codegen instead loads each operand + stores each result every iteration -- run 20,000,000 times, that is the ~1.9x. This shows the allocator turns the hot loop memory-free, the direct path to closing the gap. Known answer: 0 spills (fits in registers) AND regalloc memory traffic = 0 AND the stack baseline is large -> exit 0.