nx_bench_core.nx

nx_bench_core.nx -- B1 second stone of NISHI_PERFORMANCE_ DOCTRINE_ROADMAP.md. Shippable bench harness primitive: warmup + N-iteration measurement + reference band check + pathology classification + ETG attestation. Composes nx_etg (E1), nx_perf_pathology (B1 sibling), and nx_clock (existing monotonic timer). Why this brick is the SUBSTRATE BENCH PATTERN: every future bench (CPU matmul, memory copy, NVMe scan, network RTT, GPU kernel) calls this harness with a "workload runner" + a reference band + signals. The harness: 1. Runs N_WARMUP iterations to settle caches / JIT / TLB 2. Runs N_MEASURED iterations capturing per-iteration ns 3. Computes median, p50, p99 (sort then index for B1 first stone; reservoir sampling deferred to B3) 4. Checks if median falls inside the reference (lower, upper) band 5. Calls nx_perf_classify with the signals + in_band flag 6. Emits NxEtgEntry via nx_etg_entry_init with: probe_kind = caller-supplied (cpu/storage/net/gpu/etc.) claim_value = upper bound of the reference band measurement_value = the classified pathology id outcome = derived from pathology (OPTIMAL -> CONFIRMED, anything else -> the ETG outcome that best matches the pathology) Composes: [[NISHI_PERFORMANCE_DOCTRINE_ROADMAP]] B1 nx_etg.nx (E1 attestation) nx_perf_pathology.nx (B1 sibling classifier) nx_clock.nx (monotonic timing) [[feedback-hardware-agnostic-is-robustness]] cardinal: this harness must produce identical pathology classifications on identical inputs across ISAs. [[feedback-no-strawman-perf-comparisons]] (no claim without paired measurement; the reference band IS the paired claim) [[feedback-no-false-ok-substrate-honesty-audit]] (INCONCLUSIVE when N too small; never silent)