Add internal evaluation semantic metadata

[uriahf]

Goal

Add non-breaking internal semantic metadata for the parity work without changing current public or rendered behavior.

What this adds

Adds an internal _EvaluationMetadata representation and _build_evaluation_metadata() helper that records the semantic information that can be known from current input shapes while preserving reference_group as the compatibility grouping key.

• shared outcome/time arrays: probability keys are identifiable as models evaluated in one shared population;
• keyed outcome/time dictionaries: keys identify distinct evaluation populations, while model identity remains None because the current API does not encode it separately;
• paired/generic labels are preserved as compatibility evaluation labels rather than parsed or guessed.

Compatibility

This helper is internal and is not yet wired into rendering, reference ownership, performance tables, or public return values. Existing trace names/counts, legends, colors, statistical calculations, dataframe schemas, and APIs are unchanged.

This is deliberately preparatory groundwork for the later behavioral reference-ownership change in #368.

Tests

Adds focused tests for shared-model, keyed-population, and paired-label semantics.

Scope

No public API changes, no statistical changes, no plotting changes, no rtichoke_viz changes.

[github-actions]

PR Preview Action v1.8.1
Preview removed because the pull request was closed.
2026-08-21 11:52 UTC