Add out-of-sample check on fresh seeds: max-safety reproduces 231.4 mean / 0.17% crashes; history variants Pareto-dominate the teacher

Replace with max-safety weights (start 1 of the 6-run frontier): mean 231.2, landings 99.67%, crashes 0.25% - 23x fewer crashes than the teacher

Add Reproducibility section: 6-start spread of the crash-averse fine-tune (0.22%..6.7%); two starts Pareto-dominate the teacher

Replace with crash-averse weights: mean 239.5, landings 99.4%, crashes 0.57% (was 247.2 / 94.1% / 5.9%); update card

Define landing rate precisely (env +100 verdict) + landing_check.py

32-parameter linear LunarLander-v3 policy (mean 247.2, 60/60 seeds)

initial commit