sketch_ewma.nx

sketch_ewma.nx -- exponentially-weighted moving average (EWMA). Single-pass time-series smoother: x_new = alpha * sample + (1 - alpha) * x_old Smaller alpha = more smoothing (longer memory). alpha=1 means "no smoothing" -- track current sample only. alpha near 0 means "very long memory." PARAMETERIZATION: alpha is expressed in parts-per-million (PPM). Valid range [1, 1_000_000]. alpha_ppm=10000 = 0.01 = "Q1 weight of 1% per sample" ~ effective window ~100 samples. INTEGER ARITHMETIC (no f64 in sibling): x_new = (alpha_ppm * sample + (1e6 - alpha_ppm) * x_old) / 1e6 OVERFLOW BUDGET: |sample|, |x_old| < 2^32 to avoid overflow in alpha * sample. For |sample| up to 2^43 (~8.8e12) safe given alpha_ppm <= 1e6 (2^20). USE CASES: - SRE latency smoothing (oncall dashboards) - financial price tracking (EMA in technical analysis) - sensor smoothing (IoT temperature/humidity) LOSSLESS-LANGUAGE DISCIPLINE: EWMA's "approximation" is the systematic bias toward recent samples -- a feature, not loss. We declare envelope_kind = NX_ENV_ABS, param_a = quantization error per step (<= 1). MaturityClass = Production (deterministic recurrence, no probabilistic bounds).