nx_eqsat_vs_gcc_battery_test.nx

nx_eqsat_vs_gcc_battery_test.nx -- FAIR apples-to-apples op-count battery for the e-graph optimizer vs gcc -O3 (incumbent oracle, same box). METRIC (stated explicitly): ARITHMETIC OPERATION COUNT of the optimized straight-line computation. We count DISTINCT arithmetic nodes in the emitted DAG (op != CONST, op != VAR), i.e. each shared subexpression once -- the SAME thing one counts in gcc's asm (add/sub/imul/lea-scale/shl/ shr/sar/and/or/xor/not/neg in the function body, excluding prologue/ epilogue/reg-moves/ret). This is NOT the recursive class_cost (which double-counts shared children and would UNFAIRLY inflate our number). NON-CHERRY-PICKED: a fixed battery declared up front; every kernel reported whatever the verdict (no dropping losers). Each kernel expresses the SAME computation we hand to gcc in battery.c. Output per kernel: "<id> <emit_total> <arith_ops> <cost> <root_op> <sat>" arith_ops is the FAIR metric. cost is class_cost (context only). sat is the saturate status (4=SATURATED/fixpoint, 5=hit-budget; extraction is sound either way -- costs only decrease, so budget exhaustion is not an error).