_sprinkler_w0_throughput_bench.nx

_sprinkler_w0_throughput_bench.nx -- W0 hardened-field embedded workload baseline. THE primary W0 (hardened field embedded) workload per WORKLOAD_TARGETS.md. Sprinkler controller / soil-moisture sensing / schedule-driven irrigation -- the smallest realistic production workload the Nishi substrate must run efficiently on a Cortex-M-class core (or eventually a Tier 0 Nishi-silicon micro). Operator cardinal 2026-05-26: "sprinkler systems and sensors and all that as the hardened field use case". Inner-loop shape (per sensor sample, sub-ms deadline): raw = adc_read(soil_moisture_pin) (~1us simulated) smoothed = ewma(prev_smoothed, raw, alpha) (1 multiply + 1 shift) if smoothed < dry_threshold && schedule_allows(now): valve_open() elif smoothed > wet_threshold: valve_close() Bench: 1,000,000 sensor samples processed. No FP; everything integer (Q12 for the EWMA fraction). Reports samples/sec + per-sample latency in ns (target: sub-microsecond on bare metal). Reference numbers for silicon-feedback ROI (W0 tier): Arduino AVR @ 16 MHz, pure-C ~10K-50K samples/sec Cortex-M0+ @ 48 MHz, pure-C ~500K samples/sec Cortex-M4 @ 100 MHz, hand-tuned ~5M samples/sec This bench, pure-NishiLang on x86_64 ~100M-1B samples/sec est What this bench produces for silicon-feedback per WORKLOAD_TARGETS.md per-tier ROI math: - Per-sample latency (deterministic; the dominant W0 metric) - Per-instruction attribution (rv64im_min_sim attribution identifies whether the EWMA multiply or the schedule branch or the comparator dominates -- the answer guides silicon) Per [[feedback-honest-perf-verdict]]: this is the baseline.