Code audit: 12 critical + 14 medium issues found

[eprifti]

Critical (crash or corrupt results)

1. sa.rs/ils.rs:155,128 — Panic when k_min > 10 and k_max == 0 (usize underflow)
2. bayesian_mcmc.rs:142 — random_normal produces -infinity when u=0.0
3. bayesian_mcmc.rs:636-637 — Exponential overflow in posterior accumulation
4. gpu.rs:767 — Over-copying from staging buffer reads stale data
5. gpu.rs:687 — Panic on empty models vector
6. beam.rs:79,126 — HashSet non-deterministic iteration breaks reproducibility
7. lasso.rs:36,224 — Panic on empty features or n_alphas=0
8. individual.rs:292-330 — unwrap() on test metrics without checking Some
9. population.rs:1006,1362 — Panic/div-by-zero on empty population
10. data.rs:999,1013 — Division by zero in Bayesian Fisher
11. lib.rs:131 — Silent error discards data load Result
12. param.rs:833 — Missing n_chains in valid MCMC keys

Medium (correctness, performance, maintenance)

13. individual.rs:2582-84 — Unnecessary unsafe impl Send/Sync
14. individual.rs:1244 — Inconsistent threshold comparison (> vs >=)
15. individual.rs:1931 — Full data.X clone per permutation in MDA
16. data.rs:1175 — O(n²) in select_features
17. voting.rs:1680 — O(experts × samples²) in display
18. experiment.rs:235 — partial_cmp().unwrap() panics on NaN
19. experiment.rs:523 — assert_eq! in display crashes instead of warning
20. bayesian_mcmc.rs:919-981 — Duplicated LASSO pre-screen
21. lib.rs:125-617 — ~200 lines triplicated experiment-building