                        Changes in version 1.0.4                        

  - Apply the same Armadillo OpenMP configuration to all C++ translation
    units, including the generated Rcpp wrappers. This fixes the
    inconsistent sparse matrix type definitions reported by CRAN's GCC
    link-time optimization check.
  - Preserve explicit parallelism across cross-validation folds, with
    sequential Armadillo operations inside each fold. Estimation
    methods, numerical kernels, public interfaces and defaults are
    unchanged.

                 Changes in version 1.0.3 (2026-10-11)                  

  - Restore default prediction and coefficient extraction on current R
    versions.
  - Path objects return the full coefficient or prediction path when s
    is omitted. Numeric s selects exact fitted lambda values. For path
    objects, s = "lambda.min" selects the smallest fitted lambda;
    lambda.1se requires CV.
  - Forward num_threads and converge from cross-validation wrappers, and
    honor the supplied initial converge value in the solver. The default
    is unchanged.
  - Reject nonfinite data, constant predictors, invalid penalty
    parameters, invalid fold counts and incompatible restriction
    dimensions before fitting.
  - Replace deprecated Armadillo scalar conversion and remove obsolete
    C++11 requests. Reuse the fixed design cross-product in the
    restricted solver.
  - Update authorship, dependencies, published article citation and
    documentation.
  - Cache fixed and active-set cross-products in both solvers. Solve the
    restricted bridge and restriction right-hand sides together instead
    of repeating matrix inversions. Singular or numerically unresolved
    systems now fail explicitly; no automatic approximate solution is
    substituted.
  - Reuse CV data slices and compute prediction paths in one matrix
    multiplication. Request portable OpenMP flags where available, with
    a sequential fallback. Report the installed build capability with
    the additive rbridge_info() API. Worker failures are propagated
    after threads finish.
  - Add 36 frozen synthetic regression cases, independent
    constrained-ridge and uneven-fold CV checks, thread consistency
    tests and locally built guides.
  - Document a completed 96-shard synthetic benchmark with 48 passing
    numerical comparisons. Casewise median speed ratios are 3.14x for
    bridge, 2.59x for rbridge, 2.99x for cv.bridge and 2.57x for
    cv.rbridge against the first repaired candidate. Reported ratios are
    limited to the recorded designs and environment.
  - Preserve the historical solver, lambda grid, folds, thresholds and
    the standardized coefficient scale used for restrictions. Returned
    coefficients remain on the original predictor scale.

                 Changes in version 1.0.2 (2020-02-29)                  

  - Previous CRAN release, published on 2020-02-29.
