nx_optimize_suite_test.nx

nx_optimize_suite_test.nx -- the Nishi TEAM optimizing a SUITE autonomously (the mechanism for grinding benchmarks). For each kernel the loop: SEARCHES for a cheaper equivalent (e-graph superopt), MEASURES the cost gain, VERIFIES the candidate still evaluates equal (1:1), and the crew GOVERNS the swap -- absorbing real wins, correctly leaving already-optimal kernels alone. No hand-optimizing: the loop builds + optimizes. This is point-and-deliver in miniature. Suite: mul x K for K in {3,4,8,16,32}. K=2^k strength-reduces (mul->shl, cost 3->1, ABSORBED); K=3 has no cheaper equivalent (cost stays 3, DEFERRED -- the loop honestly banks no false win). Known answer: "absorbed=4 deferred=1 gain=8".