fix: calibrate generated risk estimates

[cr-shah]

Finding

The demo and PaySim workflows fed uncalibrated model outputs into expected-loss ranking even though the architecture and model card describe calibrated probabilities. The existing prefit helper also used an API removed by the locked scikit-learn version, and reporting could reselect a model from held-out test performance.

Fix

• select the base model only from chronological validation average precision
• fit sigmoid calibration on validation data with current and 1.5-compatible scikit-learn paths
• require reporting to use the validation-selected calibrated model
• regenerate all affected report artifacts through the project command and synchronize README metrics
• add regression checks for prefit calibration and generated-result documentation

Verification

• Ruff: passed
• pytest: 60 passed
• uv lock check offline: passed
• installed dependency compatibility: passed
• two fresh 12,000-row demo runs matched all generated JSON, CSV, and PNG artifacts byte-for-byte

This is separate from open PRs #1-#3 and does not touch their files.