feat(governance): add anomaly detection for inference outputs

[obielin]

Summary

• Adds src/climatevision/governance/anomaly_detector.py — a hybrid detector that combines an Isolation Forest fitted on rolling prediction history with a statistical fallback (z-score + IQR fences) for the cold-start case before there is enough history to fit IF.
• Persists feature history to JSONL so detection survives across processes and restarts.
• Emits anomaly reports (write_anomaly_report) for human review of flagged predictions.
• Wires the new symbols into governance/__init__.py.

Why

Sprint deliverable: "Create anomaly detection for inference inputs and outputs — flag unusual predictions for human review." Pairs with the upcoming /api/anomalies endpoint and the audit-logger PR.

Test plan

• pytest tests/test_anomaly_detector.py — 6/6 pass locally
  • feature extraction shape/range checks
  • empty-input rejection
  • statistical fallback flags an outlier after seeding normal history
  • Isolation Forest activates once history reaches min_history_for_iforest
  • history persistence round-trip across detector instances
  • JSON anomaly report writes correctly

Notes for reviewers

• Branched off develop. No overlap with PR feat(governance): add SHAP explainability for segmentation predictions #29 (SHAP, merged) or feat(governance): add regional bias audit framework for model fairness #30 (bias audit, open).
• Default contamination for IF is 0.05; tune via AnomalyDetector(contamination=...).

[femi23]

Hybrid IF + statistical fallback with rolling-history persistence is the right shape — this fills the gap I left in #41 (alert_generator) where I wanted a smarter signal than 'value > threshold'. The feature vector (mean / std / positive_fraction / entropy) is small enough that fitting IF on 50 samples is cheap, which is what we need at our throughput.

Approving.