Changes in version 0.1.0                        

First public release. The registry-dispatched facade of the development
scaffold gains its value layer: the decision-relevant estimands and the
cross-method consistency checks that make survkit a methodology rather
than a wrapper.

Breaking changes

  - survkit_tidy() no longer exponentiates a parametric fit's baseline
    parameters into a meaningless ratio, and it now labels each
    covariate effect by scale. The output gains a component column
    ("covariate" / "auxiliary") and a ratio_type column ("HR" / "TR" /
    "OR" / probit).
    
    # before: shape and scale were listed as covariates with a bogus "ratio"
    #   term  estimate  ratio
    #   scale 6.27      531.0        <- exponentiated log-scale intercept
    #   sex   0.38      1.47         <- a time ratio, unlabelled
    
    # after: auxiliaries are separated, covariates carry their ratio type
    #   term  component  estimate  ratio_type  ratio
    #   scale auxiliary  6.27      <NA>        NA
    #   sex   covariate  0.38      TR          1.47

  - survkit_curve() returns its predicted quantity in a column named
    value (was survival), because the function now predicts hazard,
    cumulative hazard and quantiles as well as survival.

New features

  - survkit_triangulate() fits the same data under several methods and
    reads their relative spread on the restricted mean (and, optionally,
    survival at a landmark) as a model-misspecification signal.
  - survkit_rmst() computes the restricted mean survival time to a
    horizon, with the backend's interval for parametric fits and the
    survival package's standard error for Cox-family and non-parametric
    fits.
  - survkit_cuminc() gives the Aalen-Johansen cumulative incidence of
    each competing event.
  - survkit_score() scores predictions with Harrell's concordance and
    the inverse-probability-of-censoring-weighted Brier score and its
    index of prediction accuracy.
  - survkit_contract() and survkit_validate_method() expose and check
    the registry conformance contract every fit closure must honour.
  - New methods in the registry: km (Kaplan-Meier / Nelson-Aalen
    non-parametric baseline) and cox_ridge (ridge-penalised Cox for the
    few-events regime).
  - survkit_curve() gains a type argument ("survival", "hazard",
    "cumhaz", "quantile").
  - as_survkit_manifest() now emits a genuine per-draw survival ensemble
    for parametric, spline and cure fits (the draws and the survival
    share one sample); Cox-family fits emit the maximum-likelihood curve
    marked draw = 0.
  - survkit() gains an on_underpowered = c("warn", "abstain") argument.
    "warn" preserves the historical behaviour (fit anyway, with a
    warning); "abstain" declines to fit an under-powered model and
    instead returns a [survkit_refusal()] -- a typed, classed token
    (c("survkit_refusal", "orchestra_refusal", "error", "condition"))
    recognisable to a leader-side gate via is_orchestra_decline(),
    rather than a bare warning() plus a fitted model regardless of the
    power gate's verdict.

Minor improvements and fixes

  - The power gate returns "unknown" for a multi-state (competing-risks)
    response rather than pooling the competing events into one
    misleading count.
  - Cox-family fits retain their model frame (model = TRUE), so survival
    curves and scoring on new data work reliably.
  - survkit_fit carries ratio_type and aux_pars so the tidier can
    separate covariate effects from baseline parameters for any backend.

Documentation

  - Vignette (Getting started with survkit) gains three figures with
    captions and interpretive prose: a Kaplan-Meier curve by age group,
    a stacked cumulative-incidence plot for the competing-risks example,
    and a "triangulation fan" plotting each method's restricted mean
    survival time against the disagreement band. It also states the
    governing equations for the events-per-variable gate, the restricted
    mean survival time and the IPCW Brier score, each with its source
    citation, and rewords the triangulation section so it no longer
    asserts that a flagged disagreement proves model misspecification
    (it can also arise from comparing different estimands, which the
    vignette now says explicitly).
  - README.md leads with the package's purpose in the first lines, and
    every code snippet is now self-contained and runnable end-to-end on
    survival::lung / survival::mgus2 (the Fine-Gray, small-sample-gate
    and scoring snippets previously referenced undefined objects).
  - Roxygen: several @titles rewritten as sentences/imperatives rather
    than noun phrases (survkit_power, survkit_fit, survkit_manifest,
    survkit_contract, survkit_cuminc, survkit_rmst, survkit_curve);
    @examples added for survkit_fit, survkit_manifest and the predict
    method on a fit (previously undocumented by example); @references
    added to survkit_power and survkit_score; governing-equation
    @details added to survkit_rmst and survkit_score.