feat: add calibration disagreement explorer

[BLCCoreStudio]

Summary

Add an evidence-level calibration disagreement explorer for TurkishEvalKit 0.9.0.

What changes

• add a typed DisagreementReport core API for scalar, pairwise, and audio calibration inputs
• expose criterion-level evaluator observations, differing evaluator pairs, score/preference gaps, and human evidence notes
• keep pairwise holistic preference/strength differences separate from criterion evidence
• expose unmatched audio annotations plus matched timestamp/severity variance using the saved calibration tolerance
• add GET /api/calibrations/<artifact>/disagreements
• derive historical attribution from the saved calibration source snapshot and immutable source evaluations
• return an explicit conflict if a referenced source evaluation is no longer available rather than inventing partial evidence
• add a browser drill-down with stale-request protection when rapidly switching history items
• add public Python exports, dedicated documentation, and a feature-specific CI gate
• align package/public version metadata at 0.9.0

Trust boundary

The explorer is a derived read-time projection. It does not persist evaluator rankings, disagreement priority, or a second calibration truth. It does not decide which evaluator is correct and makes no external AI calls.

Validation before PR

The push CI quality matrix is green on Python 3.11, 3.12, and 3.13 for Ruff, strict mypy, full pytest/coverage, wheel package-data smoke tests, and installed CLI smoke tests. Live localhost/browser regression jobs are also part of the same CI run and will be checked before merge.