Added coe.aggregate = "max.robust" as the default COE ensemble aggregation for new fits, with additional protection against repeated maximum tree contributions. The correction can also apply at coe.trim = 0; trimmed.mean, median, and mean remain available.
Extended OOB trim selection and fitted hazard configuration to max.robust. Prediction reuses the fitted aggregation policy and selected trim, while restore recomputes the trim selection from OOB risk.
Standardized coe.trim = 0.5 as an explicit exact-median endpoint for both max.robust and trimmed.mean, including fixed, adaptively selected, and inherited prediction settings, with consistent risk evaluation and selected-aggregation metadata.
Updated auct.rhf() to use the fitted max.robust rule when reconstructing cumulative/dynamic markers and recovering incident markers from retained tree-level arrays, including repeated-maximum correction and the exact-median endpoint.
Added direct Surv(time, event) support to fitting, prediction, restore, and tree-size tuning. Ordinary right-censored data are converted internally to one interval per subject, fitted forests retain the original response map, and predict.rhf() accepts test data using the original time/event column names; the four-argument counting-process interface remains available for time-dependent covariates.
Distinguished public prediction domains by original response format across grow, restore, and outcome-bearing prediction. Right-censored, time-static fits now return hazard and CHF over the complete forest horizon; counting-process fits keep hazards NA outside supplied paths, keep CHF flat through internal gaps, and set CHF to NA after each subject's final stop.
Made the tree-size tuning preflight use the same normalized data, event-process classification, and event-supported bounds as rhf(), including when an explicit upper bound is supplied.
Corrected cumulative/dynamic auct.rhf() to reconstruct common-time markers from retained tree-level hazard/CHF arrays using the fitted ensemble aggregation rule, including Nelson–Aalen mean aggregation and the exact-median endpoint, while leaving public path-domain predictions unchanged. AUC output now records marker diagnostics and reports a clear error when required raw arrays are unavailable.
Corrected the incident-AUC tuning example to use a feasible grouped-case threshold, preventing an all-nonfinite criterion when --run-donttest examples are checked.
Added the native Conservation-of-Events (COE) terminal-node hazard estimator and robust across-tree aggregation. The default adaptive protocol selects a winsorization fraction by OOB risk; median, mean, fixed trimming, and the legacy Nelson–Aalen estimator remain available.
Corrected OOB hazard risk, cumulative-hazard integration, and case-specific trajectory construction to use exact start–stop/grid overlap.
Case-specific hazards are now NA outside the supplied (start, stop] paths, while cumulative hazards remain flat through gaps and after follow-up.
Made grow, restore, and outcome-bearing prediction share the fitted hazard configuration and return coherent risk, membership, tree-level hazard/CHF, and terminal-node U, V, and COE outputs.
Strengthened counting-process validation and event bookkeeping. Unsupported recurrent-event histories are now detected and reported explicitly.
Reworked default tree sizing and tree-size tuning to use observed event support, preserve time-grid and aggregation settings, and reproducibly refit the selected forest.
Improved auct.rhf() subject-level bootstrapping and expanded time-localized VarPro tools with pilot-forest predictor weights and bar-matrix, dot-matrix, and line displays.
Renamed the survival split rules to hazard.loglik and hazard.nelson.aalen. Importance plotting arguments now use dotted names, xvar.wt.rhf() uses the pilot/cache interface, and treesize should be numeric.
Maintenance update with bug fixes and CRAN-related cleanups.
Added explicit POSIX socket/select header includes in src/server.h for improved musl/Alpine compatibility.
Replaced unsuppressible console output in print.rhf(), print.auct.rhf(), and RHF importance helpers with suppressible message-based output.
Hardened prediction and native-interface handling, including more robust checks for variable-level metadata passed from R to compiled code.
Updated native routine registration to match the current C entry-point signatures and applied minor build/source cleanups.