Add canonical PerformanceTableSpec producers

[uriahf]

Summary

• add internal static and time-dependent canonical PerformanceTableSpec builders
• map existing calculated performance quantities to canonical long-form metric values
• preserve semantic evaluation metadata, model-known/model-unknown population identity, operating points, horizons, and heuristic context
• keep existing Great Tables / Reactable production rendering unchanged
• add focused semantic tests for identity, deterministic evaluation IDs, operating points, metric IDs, zero/null handling, time context, and equal-valued distinct evaluations

Architecture

Boundary remains:

prepare_performance_data[_times]() → _build_evaluation_metadata() → canonical PerformanceTableSpec

No statistics are recomputed in the builders. reference_group is used only to join calculated rows to semantic metadata. No seriesId is introduced.

Contract note

The canonical TS contract currently permits exactly the metric IDs emitted here. Existing production columns that are not canonical metric IDs (real_positives, real_negatives, predicted_negatives, n) are intentionally not emitted.

Validation

CI requested. Local execution was not available in this environment, so Ruff/format/ty/pytest/build results will be taken from GitHub Actions.

[github-actions]

PR Preview Action v1.8.1
Preview removed because the pull request was closed.
2026-08-24 09:32 UTC