nx_benchmark_harness.nx

nx_benchmark_harness.nx -- head-to-head perf verdict primitive. Substrate-level enforcer of the cardinal feedback-honest-perf- verdict-no-aspirational-claims. Every NishiLang vs incumbent benchmark emits a SEALED-ENUM verdict per axis: WIN / LOSE / TIE / UNMEASURABLE. No "competitive with" weasel words. Every LOSE row MUST be paired with a named-improvement string the caller fills in. IDEA-PROVENANCE (per the language-design-research discipline: learn from, never copy): - JMH (OpenJDK) -- multi-warmup, multi-fork, GC-quiet - Criterion.rs -- bootstrap CI, outlier detection (MAD) - Google Benchmark -- statistical-rigor patterns - Hyperfine -- multi-trial CLI command timing Every idea was re-derived from the published papers + project READMEs; no source code was incorporated. See genealogy_id for paper citations. Six measurable axes per benchmark run: 1. THROUGHPUT ops/sec or items/sec 2. LATENCY_P50 median single-op time 3. LATENCY_P99 99th-percentile single-op time (tail latency) 4. MEMORY_PEAK peak bytes during run 5. DETERMINISM same-input-same-output across N trials (Q10) 6. PORTABILITY runs without modification across target archs (Q10) Each axis emits a per-axis sealed-enum verdict: NX_BENCH_AXIS_WIN challenger >= incumbent * 1.01 (1% delta floor) NX_BENCH_AXIS_TIE within 1% either direction NX_BENCH_AXIS_LOSE challenger < incumbent * 0.99 NX_BENCH_AXIS_UNMEASURABLE measurement not available this run Composite verdict: NX_BENCH_DECISIVE_WIN >= 4 of 6 axes WIN, zero LOSE NX_BENCH_WIN majority WIN, fewer LOSE NX_BENCH_TIE majority TIE or balanced WIN/LOSE NX_BENCH_LOSE majority LOSE NX_BENCH_DECISIVE_LOSE >= 4 of 6 axes LOSE NX_BENCH_INCONCLUSIVE too many UNMEASURABLE to call Why 1% floor: per cardinal feedback-one-percent-minimum-delta-substrate