Characterize static evaluation semantics

[uriahf]

Goal

Implement Stage 1A of the cross-language semantic audit as characterization tests only.

Coverage

Adds focused tests for:

• one model × one population
• multiple models × the same population
• the same model/prediction rule across populations with different prevalence
• distinct populations with equal prevalence
• explicitly paired model-population inputs
• calibration grouping and the global identity line
• performance-table input semantics
• ROC, precision-recall, gains, lift, decision, and interventions-avoided plotted grouping/reference ownership
• current color behavior for the single-model case

The tests use the Stage 0 vocabulary while deliberately recording the current R representation: model-labelled shared-outcome inputs use model, paired/multi-population inputs use population, and reference ownership changes by output.

Scope

Test-only. No statistical calculations, public APIs, production behavior, plotting code, package metadata, generated documentation, or rtichoke_viz are changed.

Any surprising current behavior is characterized rather than fixed in this PR.

Paired PR

Python: uriahf/rtichoke_python#365