nx_hazard_lib.nx

nx_hazard_lib.nx -- THE WEAR-AND-TEAR CORE. Weibull survival in pure integer fixed point. WHY THIS EXISTS (measured, not assumed, 2026-08-08): nx_flip_score decides per component with `margin = lift - repair`. Verified over all 8 parts of the live FOCUS13A deal: EVERY margin is exactly lift-minus-repair. The findings plane carries a `cond` (condition 0-100) column -- and it appears in ZERO decisions. It prints a number, it reads as evidence, and it drives nothing. There was no mileage, no age, no failure probability and no expected cost anywhere in the engine. `nx_capsearch` over 960 registered tools returned no survival/hazard organ. So this is net-new, and it is the half the operator named: "the fail points and wear and tear". THE MODEL: two-parameter Weibull, S(t) = exp(-(t/eta)^beta). beta (SHAPE) is literally the wear regime, and that is why this model and not a flat rate: beta < 1 infant mortality -- a DEFECT/recall class; risk FALLS with miles beta ~ 1 memoryless random failure; miles tell you nothing beta > 1 WEAR-OUT; risk RISES with miles <- "wear and tear", as one number eta (SCALE) characteristic life; S(eta) = 1/e, i.e. 63.2% have failed by eta. THE QUANTITY THAT ACTUALLY PRICES A DEAL is not P(fail ever) but the CONDITIONAL probability of failing during the holding window, given the part already survived to today's odometer: P(fail in (t0,t1] | survived t0) = 1 - S(t1)/S(t0) = 1 - exp( (t0/eta)^beta - (t1/eta)^beta ) Measured consequence on a real fitted curve (beta=2.5, eta=150000): the SAME part with the SAME repair cost carries 335 bps of risk over a 12k-mile hold at 40k miles, and 1427 bps at 120k -- 4.3x. An engine without this prices both at zero. ESTIMATOR: median-rank regression (the Weibull probability plot), Benard's approximation F_i = (i-0.3)/(n+0.4), then least squares of ln(-ln(1-F_i)) on ln(t_i). Slope IS beta. Chosen over MLE deliberately: it is exact-arithmetic-friendly, it degrades honestly on small samples, and it is THE reliability-engineering standard, so a third party can check our number. ASSET-AGNOSTIC BY CONSTRUCTION -- there is no "car" in this file. `t` is any monotone usage/age measure: miles for a car, months-in-force for a bond, cycles for a battery, days-held for a position. Same curve, same estimator. That is what makes the omni ask reachable. PRECISION: Q20 fixed point (HZ_FP = 1048576 = 1.0), i64 throughout. Measured against Python ground truth: beta recovered to <1 bps, eta to <0.3%. Callers MUST treat eta as +/-1%. license_tier: ORIGINAL No hw writes (Rule 26).