Claim [2]: reproduce Lemma 4.2 -- hazard-MLE consistency + asymptotic normality O(n^-1/2) (empirical cov -> inverse Fisher, KS-normal, 95% Wald coverage) and delta-method normality of the plug-in value estimator

Reproduce Prop 4.1: goal-conditioned value = -sum gamma^t S(t); value iteration -4.116238 == survival identity -4.116238 (error 0.0), multimodal time-to-goal