nx_daemon_bench.nx

nx_daemon_bench.nx -- continuous, low-resource, on-device benchmark daemon. Substrate-native equivalent of Criterion.rs / JMH / Google Benchmark / Phoronix Test Suite, baked into the language itself and explicitly designed for Tier-0 (MCU/sensor) operation. license_tier: ORIGINAL MISSION (per FLoC-Olympics + monetization cardinal + sovereign- stack roadmap): NishiLang must compete against and beat every modern continuous-benchmark system, AND the same primitive must run on a sensor with kilobytes of RAM. Sensors that audit themselves at <1% CPU duty cycle are a new product category. DESIGN PRINCIPLES: 1. FIXED-MEMORY STATE. Single mmap at init; never grows. Each derived statistic uses Welford's online algorithm (mean + variance in O(1) memory regardless of cycle count). No sample log, no histogram bins, no growing record ring. Sensor-grade footprint: under 1 KiB of state. 2. DUTY-CYCLE BUDGET. Caller declares max ms/cycle, max bytes, and min_sleep_ms between cycles. Default 100 ms work + 10000 ms sleep -> <1% CPU. Daemon self-polices: cycles exceeding the wall-time budget emit OVERBUDGET and increment a counter. 3. SEALED-ENUM STATUS. Eight states, no weasel intermediates, validated by predicate. Cardinal: feedback-honest-perf-verdict-no-aspirational-claims. 4. REGRESSION DETECTION. Per-axis EMA baseline + Welford variance -> Z-score in Q10. |Z| > 3-sigma -> DEGRADED; |Z| > 5-sigma -> REGRESSION (cardinal requires named_improvement string from caller). 5. TIER-AWARE SKIP. Caller passes the current hardware tier; if below cfg.tier_floor the cycle is skipped (SLEEPING). Cardinal: scale-agnostic-substrate. 6. CARDINAL-COMPLIANCE GATE. Refuses to declare a REGRESSION