nx_bench_harness_test.nx

nx_bench_harness_test.nx -- KAT for the leaderboard engine. Native sovereign lane; exit 0 = pass, N = assertion N failed. Proves three things on toy-but-exact data: 1. an IDEAL ordering scores the exact reference values (nDCG=MAP=MRR= canonical=1.0; primary precision + S-recall = the closed-form fractions) 2. a JUNK-FIRST ordering (reaction video + listicle on top) is penalized 3. nx_bench_nishi_rank re-orders the SAME documents into a measured S-class EXCEED over the junk ordering (nx_bench_exceeds == 1)