Start v2 safety hardening

fix: wilcoxon sys.path for condition_a import, zero_method=pratt for binary scores (Day 15)

fix: Wilcoxon uses stub guardrail — tests retrieval quality on all 50 prompts not 13 (Day 15)

fix: threshold sweep — calibrate DeBERTa guardrail threshold, patch eval imports (Day 14)

feat: eval scripts — BM25 baseline, adversarial probes, BERTScore, RAGAS, Wilcoxon (Day 14)