Characterize time-dependent evaluation semantics

[uriahf]

Goal

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

Coverage

Adds focused tests for time-dependent semantics across multiple horizons:

• one model × one population
• multiple models × the same population
• population-labelled evaluations with different event prevalence
• population-labelled evaluations with equal prevalence
• explicitly paired model-population inputs
• ROC, precision-recall, gains, lift, decision, and interventions-avoided grouping and reference ownership
• calibration identity-line and series behavior
• plotted-series labels, colors, legend participation, and cutoff-marker behavior
• performance-data and performance-table grouping by reference_group × fixed_time_horizon

Characterized current behavior

The time-dependent input/output contract does not carry separate model and population dimensions. Multiple keyed evaluations and explicit model-population pairs remain generic reference_groups.

Reference-line scope is currently inferred from whether horizon-specific event risks differ across reference_groups. As a result:

• multiple models sharing one outcome population get shared references at each horizon;
• distinct populations with different horizon-specific risks get population-owned references;
• distinct populations with equal horizon-specific risk remain separate plotted series but their prevalence-dependent references collapse to shared references;
• reference scope can switch from shared at one horizon to population-specific at another if the population risks are equal at the first horizon and diverge later.

This PR records those semantics rather than changing them.

Scope

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

Verification

The final diff contains one added test file only. Python package CI passes on 3.12, 3.13, and 3.14, including Ruff lint/format, type checking, package build, and the full 169-test suite.

[github-actions]

PR Preview Action v1.8.1
Preview removed because the pull request was closed.
2026-08-21 11:07 UTC