Changes in version 0.4.0                        

Sparse and survival SDR

  - Added fit_risdr_sparse() for joint structural-dimension and sparsity
    selection with continuous, categorical, and right-censored survival
    outcomes.
  - Added fit_risdr_dual() and fit_risdr_sparse_realdata() for
    high-dimensional predictors, with selected-direction mapping and
    variable-ranking records.
  - Added censoring-status/time slicing, canonical pairwise-slice DR
    kernels, coordinate-level C1F adaptive weights, and complete
    information-criterion grids.
  - Retained the frozen thesis convention: Efron Cox partial likelihood,
    support_penalty = FALSE, and C1F complexity for ICOMP/CICOMP.
  - Preserved all existing 0.3.1 continuous-response interfaces and
    defaults. The new engine is explicitly selected through the new
    entry points.
  - Added executable survival examples, a migration vignette, provenance
    records, and regression checks against both the frozen and existing
    implementations.

                     Changes in version 0.3.1.9000                      

Post-release development

  - Updated package, citation, installation, CRAN, and Zenodo metadata
    following the first CRAN release.

                 Changes in version 0.3.1 (2026-07-28)                  

First CRAN release

  - Published version 0.3.1 on CRAN on 28 July 2026 and archived the
    release on Zenodo.
  - Added executable examples for the principal simulation, covariance,
    sufficient dimension reduction, prediction, slicing, structural
    dimension, and subspace-assessment workflows.
  - Hardened dependency, metadata, URL, spelling, and citation records
    for the first CRAN submission.
  - Declared both knitr and rmarkdown as vignette builders so that
    vignettes rebuild correctly when suggested packages are not
    pre-installed.
  - Added a package spelling dictionary and R-hub validation workflow.
  - Corrected structural-dimension examples to request an admissible
    maximum dimension directly, avoiding expected dimension-reduction
    warnings.
  - Excluded the repository-level CITATION.cff file from source-package
    builds while retaining inst/CITATION for the installed R package.
  - Retained the Zenodo DOI as the permanent identifier for the archived
    version 0.3.0 release and made the installed-package citation report
    the installed package version.
  - Completed local, GitHub Actions, pkgdown, coverage, URL, spelling,
    and selected R-hub checks successfully.

                        Changes in version 0.3.0                        

Package interface

  - Promoted repeated predictive cross-validation and complexity-aware
    cross-validation to the exported API.
  - Added explicit cov_args, stabilization_args, and sdr_args controls
    to prevent component arguments from leaking into unrelated
    functions.
  - Improved prediction for one-row inputs, reordered named columns, and
    alternative structural dimensions.
  - Restricted the real-data workflow to its verified
    continuous-response scope.

Numerical and validation changes

  - Added ridge and Ledoit-Wolf covariance methods to central method
    validation.
  - Replaced the observation-wise Ledoit-Wolf accumulation loop with an
    algebraically equivalent vectorised calculation.
  - Vectorised the principal Hessian directions kernel calculation.
  - Strengthened integer, missing-value, covariance, slice,
    survival-indicator, and resampling validation.
  - Corrected rho = 0 handling for AR(1) simulation and guarded
    sparse-basis generation when the predictor dimension is small.
  - Added a user-supplied true-basis option and ensured that simulated
    training and test samples share the same central subspace.
  - Made projection matrices robust to non-orthonormal full-rank bases.
  - Return NA with a warning when MAPE is undefined because an observed
    value is zero.

Documentation and release engineering

  - Added five vignettes, a structured pkgdown configuration, a
    professional README, contribution guidance, issue templates, and
    release documentation.
  - Expanded the testthat suite across validation, covariance
    estimation, SDR kernels, prediction, dimension selection,
    resampling, simulation, plotting, and real-data helpers.
  - Added multi-platform R CMD check, test coverage, and pkgdown
    workflows.
  - Added selected supplied thesis outputs under inst/extdata as
    provenance fixtures. Legacy simulation prediction outputs are
    explicitly marked for targeted rerun because the original script
    regenerated the true basis in the test sample.
  - Added the supplied aligned EPI training and test matrices, with
    dimensions, column alignment, and SHA-256 provenance recorded in
    inst/extdata/README.md.
  - Added a configurable corrected reproduction script for Simulations
    A, B1, and B2. Corrected files use distinct names and do not
    overwrite legacy outputs.

Compatibility

  - The principal user-facing function names from the supplied baseline
    are retained.
  - Version metadata now records the agreed development release
    number, 0.3.0.