nx_eqsat_race_bench.nx

nx_eqsat_race_bench.nx -- RACE TIMING HARNESS vs egg 0.11.0 on the SAME task. THE RACE: egg's /tmp/race_egg/examples/nx_race.rs timed E2 "(* x 8)" : parse-string -> RecExpr -> Runner -> saturate -> Extractor::find_best -> to_string == (<< x 3) 100000 iterations = 783465 us total = 7.835 us/op (LatCost model). OUR analogue, ONE timed iteration (the same end-to-end pipeline, no proxy): nx_eqsat_init (fresh e-graph -- egg's RecExpr alloc) add_var x, add_const 8, add_binary(MUL, x, 8) (BUILD the (* x 8) input -- egg's parse) nx_eqsat_saturate(g, 16) (egg's Runner::run to saturation) nx_eqsat_recompute_best(g) (egg's Extractor bottom-up cost pass) nx_eqsat_extract_best_node(mul) (egg's find_best for the mul e-class) read the root op of that node (egg's to_string root-symbol read) CONFIRM the extracted op == NX_EQ_OP_SHL (11) and shift const == 3. Everything that egg counted (parse->saturate->extract->stringify-root) is inside OUR timed loop too, so the us/op is comparable 1:1. We do NOT include the membership PROOF/provenance machinery (egg has none) -- only the same optimize-and-read-the-answer work egg performed. FAIL LOUD: if the extracted representative is not (shl x 3) the harness sys_exit(nonzero) BEFORE printing timing -- a fast wrong answer is not a win. Output (one line): RESULT_OP=11 SHIFT=3 ITERS=100000 TOTAL_US=<t> US_PER_OP=<u>