Changes in version 0.9.4.9002 - Fixed smooth-specification replay by requiring exact field-name lookup, so select, selection, selection_penalty, lambda, and lambda_initial cannot be confused by R list partial matching during fitting, screening, or bootstrap refits. Bootstrap refits now also discard data-specific private smooth state before rebuilding each resampled design. - Fixed null-space alias handling to retain the full independent penalized eigenspace before dropping redundant unpenalized coordinates, while still rejecting genuine penalized aliases. - Added a dependency-free penalized-spline engine for explicitly specified smooth mean terms and as an alternative to restricted cubic splines in nonlinear screening and final fitting. - Implemented term-specific matrices and smoothing parameters, penalty diagonalization, explicit range/null-space separation, term- appropriate alias/identifiability constraints, and fixed-block profiling. Smooth penalties remain separate from elastic-net coefficient shrinkage, and smooth means and serial covariance are fitted jointly for every supported response family. - Added ordinary and cyclic P-splines, factor-by/treatment-specific and numeric- by varying-coefficient smooths, tensor products, shrinkage smooths, and dependence-component grouped-CV whole-term selection. Subject-specific random smooths fail explicitly pending redesign of sparse latent integration. - Added fold-local, dependence-component grouped smoothing optimization. Training-fold boundaries, bases, constraints, screening, and transformations are isolated from validation rows, which use stored training blueprints. Only positive-weight rows form the fold graph, so inactive bridge rows cannot join otherwise independent components; constant boundary extrapolation is the safe default for noncyclic validation and prediction. - Added smooth-specific inference and reporting based on effective and reference degrees of freedom rather than raw basis counts, plus prediction reconstruction from stored smooth metadata and extrapolation rules. Ordinary covariance and direct delta-method prediction inference are suppressed consistently after penalization or grouped whole-term selection. - Updated fit, screening, prediction, summary, and print documentation. S3 output now separates parametric coefficients, elastic-net penalties, and smoothing parameters without repeating basis-coordinate or warning blocks. - Expanded numerical, statistical, regression, API, documentation, and output- contract tests for the revised model paths while retaining native compressed- design diagnostics, exact-Gaussian fitting, reuse, and deferred fold Hessians. Changes in version 0.9.4.9001 - Replaced smoke-only coverage with numerical, statistical, regression, API, documentation, and output-contract tests for the revised model paths. - Corrected latent-state, covariance, numerical-stabilization, penalized- inference, dependence-resampling, covariance-ranking, and prediction behavior, including nested and crossed designs. - Added native compressed-design rank diagnostics, concurrent optimizer starts, prepared-model and complete-start reuse, and deferred Hessian construction while retaining the exact Gaussian path. - Added .use_data(), which returns a reproducible 50-subject, three-occasion Gaussian panel with three predictors. - Reworked function examples in documentation, and made S3 print and summary displays concise and non-duplicative. Changes in version 0.9.4 - Documentation and submission. Given that CRAN submissions is offline from Aug 5, 2026 to Aug 19, 2026 (CRAN team vacation and maintenance work), the MEMWAS version 0.9.4 cannot be submitted to the CRAN. I keep improving the package. Changes in version 0.9.3.9002 - Separated the internal numerical estimator from full public fit-object construction. Cross-validation folds now use a lightweight numerical result and bypass Hessian/Wald construction, residual tables, effective-degree-of- freedom traces, information-criterion construction, covariance-component rank diagnostics, and full public S3 fit assembly. - Deferred covariance-component Jacobian rank diagnostics for cross-validation folds. The selected full-data fit continues to run the complete diagnostic and inferential construction path. - Changed fold-local nonlinear screening to run once per fold, with knots, selected formulas, fold-specific bootstrap seeds derived from supplied screening seeds, and optional reuse of the unpenalized baseline solely as an optimizer start. - Added validation-side caching for fixed-, random-, and serial-effect design objects, response encodings, offsets, weights, and fold-specific spline bases. - Converted every native call to a directly registered namespace symbol, disabled dynamic symbol lookup, and enabled R_forceSymbols() for stricter native interface checking. Changes in version 0.9.3.9001 - Refactored the native serial-contribution calculation to eliminate the GCC 16 link-time -Wstringop-overflow diagnostic arising from an inlined std::string assignment. The calculation now uses fixed boundary codes and R-owned, exact-sized diagnostic vectors. - Validated expOU driver variances before numerical output allocation and kept the serial-contribution error path free of live dynamic C++ containers. - Added event-aligned boundary_codes alongside the retained scalar boundary_code; mixed causes are summarized as multiple_serial_boundary_causes. - Added native regression tests for finite AR/expOU calculations, non-finite inputs, contribution and exponential overflow, invalid expOU variance, mixed boundary causes, dimension validation, and empty matrices. - Fixed issues regarding the tests of memory access errors, on x86_64 Linux (... For {gcc,clang}-ASAN this may also report errors...) Wishing Wan Fang Hospital great success in the NHQA competition (Taiwan) on Aug. 10, 2026. Changes in version 0.9.3 (2026-08-08) - Newly released package.