Support calibration under independent censoring

[uriahf]

What changed

• supports censoring_heuristic="adjusted" when censored observations are present
• estimates discrete calibration points with within-bin Aalen–Johansen risks (equivalent to 1 - KM without competing events)
• estimates smooth calibration curves from leave-one-out Aalen–Johansen pseudo-observations
• applies competing-event heuristics only to events on or before the calibration horizon
• keeps competing events after the horizon in the event history so pre-horizon risk sets remain correct
• keeps polarstate and pseudostate unchanged
• avoids pseudo-observation computation for discrete-only curves
• fixes single-model time-calibration trace naming and keeps tied predictions in the same calibration group

Why

Time-dependent calibration previously rejected adjusted censoring whenever censored observations were present. Under independent censoring within the prediction groups/range, Aalen–Johansen/Kaplan–Meier estimates provide the appropriate observed risks, while AJ pseudo-observations support smoothing.

pseudostate currently requires Python 3.13 while rtichoke supports Python 3.9+, so this PR performs the same leave-one-out AJ calculation internally using the existing polarstate dependency. The adapter can be replaced later if package version constraints align.

Validation

• full suite: 36 passed
• discrete survival calibration matches lifelines.KaplanMeierFitter
• discrete competing-risk calibration matches lifelines.AalenJohansenFitter
• individual pseudo-observations match independently recomputed leave-one-out lifelines AJ estimates
• regression coverage confirms competing events after the horizon are retained
• ruff check src tests: passed
• formatting check for changed files: passed
• targeted ty check for changed files: passed

[github-actions]

PR Preview Action v1.8.1
Preview removed because the pull request was closed.
2026-08-12 12:38 UTC