Moving to the survalis verb interface

survalis 1.3.0 introduced a seven-verb interface. It keeps every prior function working – nothing you have written breaks. The granular helpers are now marked internal in the manual and print a one-time note per session; they are scheduled to become unexported in a future release. This vignette shows the recommended replacements.

The mapping

Granular (still works)	Verb interface (recommended)
`fit_ranger(f, data, num.trees = 500)`	`fit(f, data, "ranger", spec = list(num.trees = 500))`
`predict_ranger(m, nd, times)`	`predict(m, nd, times)`
`predict_*()` then `survmat_to_rmst()`	`predict(m, nd, times, type = "rmst", tau = ...)`
`predict_*()` then `survmat_to_quantile()`	`predict(m, nd, times, type = "median")` / `type = "quantile"`
`cv_survlearner(f, data, fit_ranger, predict_ranger, times, folds = 5)`	`evaluate(f, data, "ranger", times = ..., resampling = cv(5))`
`score_survmodel(m, times, metrics)`	`evaluate(m, times = ..., resampling = cv())`
`tune_ranger(f, data, times, param_grid, refit_best = TRUE)`	`tune(f, data, "ranger", grid = param_grid, times = ...)`
`benchmark(f, data, learners, times)`	`compare(f, data, models = learners, times = ...)`
`best_survlearner(bench, metric)`	`print()` a `survalis_compare` (best is flagged)
`compute_pdp(m, data, "karno", times)`	`interpret(m, "pdp", feature = "karno", times = ...)`
`compute_shap(m, nd, data, times)`	`interpret(m, "shap", newdata = nd, times = ...)`
`compute_varimp(m, times)`	`interpret(m, "varimp", times = ...)`
`plot_pdp(res, "karno")`	`plot()` a `survalis_interpret`
`estimate_rmst(m, data, tau, times, trt_col = "trt")`	`estimate(f, data, "<model>", treatment = "trt", tau = ...)`

Side-by-side

library(survalis)
f <- Surv(time, status) ~ age + karno + celltype

# 1.x
m1 <- fit_ranger(f, veteran, num.trees = 200)
p1 <- predict_ranger(m1, veteran[1:3, ], times = c(90, 180))

# 2.0
m2 <- fit(f, veteran, "ranger", spec = list(num.trees = 200), seed = 1)
p2 <- predict(m2, veteran[1:3, ], times = c(90, 180))

round(as.matrix(p2), 3)
#>       t=90 t=180
#> [1,] 0.404 0.269
#> [2,] 0.857 0.484
#> [3,] 0.649 0.435

What changed on purpose

predict() always returns a survmat for the curve types: an n x length(times) matrix with the grid in attr(x, "times"). The "t=<time>" column names are still there, so the *_survmat() metric helpers keep accepting it.
times is optional. When omitted, default_times() builds a grid over the observed event times and records it on the result (res$times), so a run is reproducible.
evaluate() / compare() take a resampling object (cv(), holdout(), group_cv(), bootstrap()) instead of a bare folds integer, and compare() evaluates every model on the same folds.
fit() engine arguments go in spec, not ....