Fix time-dependent gains reference ownership

[uriahf]

Goal

Finish the production-semantic parity stage by aligning time-dependent gains with the stable population-based reference ownership introduced in #368.

Problem

create_gains_curve_times() still used a gains-specific reference replacement path that inferred whether there were multiple populations from horizon-specific event-risk data. Distinct keyed populations with equal event risk could therefore collapse to one perfect-model reference at that horizon.

Change

Route the public time-dependent gains path through the shared population-aware time reference helper already used by ROC, precision-recall, lift, decision curves, and interventions avoided.

This preserves the existing gains formulas and public API while making reference ownership depend on stable population identity rather than numerical event-risk equality.

Tests

Add public regression coverage verifying that:

• distinct keyed populations with equal event risk retain distinct population-owned perfect-model gains references;
• multiple models evaluated against shared outcome/time arrays continue to share one population-level perfect-model reference.

Scope

Python only. No statistical formula changes, no public API changes, no R changes, and no rtichoke_viz changes.

CI note

The initial run stopped at Ruff because the refactor left four stale imports in gains.py; those are cleanup-only and do not affect the semantic change.

[github-actions]

PR Preview Action v1.8.1
Preview removed because the pull request was closed.
2026-08-21 14:55 UTC