Unsupervised anomaly score for one country (experimental)

IsolationForest second-opinion model (CenDTect-inspired). Lower anomaly_score = more anomalous. is_anomaly=true means the country-day is in the most-anomalous 26% of training distribution. EXPERIMENTAL: AUC vs supervised labels = 0.489 (below 0.65 promote floor). Each response carries passed_promote_floor=false for self-disclosure. Use the DISAGREEMENT signal (unsup_only countries) as a triage queue, not as a verdict.