nx_eqsat_congruence_bench.nx

nx_eqsat_congruence_bench.nx -- NON-TOY CONGRUENCE-HEAVY race vs egg 0.11.0. THE RACE (engine-only, SAME expression both sides): egg target = /tmp/race_egg/examples/nx_cong.rs, the congruence-heavy graph: base a = (* x 8), b = (<< x 3) -- mul-pow2 merges a~b N pairs la_i = (- a c_i), lb_i = (- b c_i) -- become CONGRUENT after a~b root = (+ (+ ... (+ 0 (+ la_0 lb_0)) ...) (+ la_{N-1} lb_{N-1})) so after the single base merge, the engine's congruence closure must propagate N parent merges (la_i ~ lb_i) -- the exact reachability work egg's deferred rebuild does and the pre-MEET nx_eqsat could NOT do. This is NOT the toy (* x 8) single-rewrite race; congruence FIRES N times per iteration. OUR side, per timed iteration (the SAME end-to-end optimize pipeline egg ran, no proxy, no membership/provenance machinery egg lacks): nx_eqsat_init (fresh e-graph) enable_meet (egg hashcons + congruence rebuild ON) build a, b, N parent pairs, root (egg's RecExpr build) nx_eqsat_saturate (egg's Runner::run -- rules + rebuild to fixpoint) nx_eqsat_recompute_best (egg's Extractor bottom-up cost) nx_eqsat_extract_best_node(root) (egg's find_best) FAIL LOUD: before timing, one verified iteration asserts congruence actually fired -- find(la_i)==find(lb_i) for ALL i (== egg's internal congruence). A fast wrong answer (congruence NOT propagated) loses, exit nonzero. Output (one line): NX_CONG_ENGINE_ONLY N=<n> ITERS=<it> CONG_MERGES=<n> TOTAL_US=<t> US_PER_OP_NS=<u> (US_PER_OP_NS = total_us*1000/iters = nanoseconds/op; divide by 1000 for us/op.) SOVEREIGN: no .sh; runs on the pinned NishiLang compiler. license_tier: ORIGINAL