[ability-spec]

Summary

accuracy_in_time() counts only observed event causes when validating prediction arrays, although the cause axis must follow the response's declared states. With states ("relapse", "death") and only death events (code 2), it rejects valid two-cause predictions. A compressed one-cause array is instead accepted and can return an incorrect accuracy.

Use Surv.states when available to preserve cause codes across subsets, retain the existing single-event fallback, and clarify the docstring. Add five regression cases for missing first/middle/last causes, all-censored data, and rejection of compressed cause axes.

Validation

• All five new cases fail on the original implementation.
• Metrics and Outcome suites: 146 passed.
• Full Python suite: 2,367 passed, 16 skipped, 5 warnings.
• Ruff lint/format and Pyright passed.

The full suite ran with two workers and coverage from python/. Existing skips include optional great-tables visualizations and dependency-specific forest paths. No estimator formulas or R-generated fixtures changed.

Checklist

• Python checks passed: Ruff, Pyright, and pytest.
• Regression tests added.
• Docstring updated.

R-parity fixture changes are not applicable: this fixes cause-axis validation without changing estimator formulas.

[rich-iannone]

Great improvement, thanks!