                 Changes in version 1.2.0 (2026-09-08)                  

  - Replaced all mlr3pipelines graph and PipeOp usage with an internal
    mlr3 regression learner providing fitted preprocessing, three-fold
    out-of-fold stacking, optional base-learner selection, and averaging
    or learned stacking. The package no longer needs mlr3pipelines to
    install, load, train or predict.
  - Removed the mlr3viz, visNetwork, tictoc, forcats, readr, lubridate,
    purrr and Metrics dependencies. Mapping, timing, regression metrics,
    categorical processing and pipeline diagrams now use internal/base R
    code. R6 is now declared directly for the internal mlr3 learner.
  - Moved mlr3filters to Suggests; correlation and variance are
    internal. Specialized filters still use their original optional
    backends, preserving their algorithms. Removed unused direct
    suggestions of mlr3misc and nloptr.
  - Preserved the sense() arguments and result element names. plot is
    now a sense_pipeline object, drawn with plot(result$plot), rather
    than an HTML widget. time_log is a base R difftime. java_mem is
    accepted but unused. Stored model internals no longer have the
    GraphLearner/PipeOp layout.
  - Fixed deployment predictions to use a model refitted on all supplied
    rows, independently of outer evaluation. Fit preprocessing and
    feature selection inside every stacking fold; perform benchmarking
    inside each ensemble fit rather than selecting algorithms using the
    full data before outer evaluation. Benchmark scores are now
    out-of-fold scores and respect the measure's direction.
  - Added handling for unseen categories, constant/all-missing numeric
    columns, stable prediction schemas, finite input validation and
    reproducible fitting without changing the caller's random state or
    Java options.
  - Bound nearest-neighbour tuning by the smallest training split, avoid
    invalid zero SVM nu/eta candidates, and let ranger choose mtry for
    each encoded schema. Random search is capped by both budget and
    n_evals; one-candidate batches prevent overshooting the evaluation
    cap. Time limits are checked between trials.
  - Added base R regression and integration tests and a runnable
    migration vignette describing the architecture, optional
    dependencies and compatibility changes.
  - Converted the migration vignette from PDF/LaTeX (.Rnw) to a knitr
    HTML vignette (.Rhtml), so building and checking the package no
    longer requires a LaTeX installation or Ghostscript.

                 Changes in version 1.1.0 (2024-06-19)                  

  - Adapted to the new Paradox syntax.

                 Changes in version 1.0.0 (2021-09-06)                  

  - Added a NEWS.md file to track changes to the package.
