feat(governance): add calibration metrics for segmentation confidence

[Hopelynconsult]

Summary

• Adds governance/calibration.py with ECE, MCE, Brier score, and reliability-bin computation for binary segmentation outputs.
• Returns a CalibrationReport dataclass that serialises to JSON, ready to be attached alongside the existing Mitchell-style model cards (feat(governance): add automated model card generator #37) and consumed by the release CI gate (feat(governance): add release CI gate for metrics, fairness, and security #39).
• Pure numpy at evaluation time — no torch dependency at gate-check time.

Why calibration matters here

ClimateVision dispatches NGO alerts based on confidence thresholds. A model that reports 0.9 confidence should be correct ~90% of the time; if it isn't, every threshold downstream is silently wrong, producing either missed events (deforestation, flood) or false alarms that erode NGO trust. We currently track accuracy/IoU/F1 but never measure whether the confidence we surface is meaningful. This closes that gap with the standard reliability-diagram metrics.

What's in the PR

• ReliabilityBin and CalibrationReport dataclasses (JSON-serialisable, mirrors the style of anomaly_detector.PredictionFeatures / model_card.ModelCard).
• evaluate_calibration() one-shot entrypoint returning a populated report.
• expected_calibration_error() (support-weighted), maximum_calibration_error() (worst-bin), brier_score().
• write_calibration_report() for persistence next to model cards in outputs/.
• is_well_calibrated() helper with a 5% ECE default — wires straight into the release CI gate's threshold pattern.
• 12 unit tests covering perfectly-calibrated and overconfident synthetic distributions, MCE ≥ ECE invariant, Brier extremes, full input validation (probability range, binary targets, shape match), and round-trip JSON serialisation.

Follow-ups (separate PRs, not in scope here)

• Wire evaluate_calibration into scripts/generate_model_card.py so every model card includes a calibration block.
• Add an ece threshold to scripts/governance_ci_gate.py so miscalibrated releases fail the CI gate.

Test plan

• pytest tests/test_calibration.py -q → 12 passed
• Reviewer: confirm threshold defaults (ece_threshold=0.05, n_bins=15) match what we want for the release gate.

🤖 Generated with Claude Code