######################## version 1.1 (2016-01-24) ######################## Initial release, the sun shines and sommer has arrived! ######################## version 1.2 (2016-02-07) ######################## + 'fdr' function bug was fixed + addition of the 'randef' function + addition of the converter 'atcg1234' function + names in the blup's or random effects added + zero-boundary constraint added to Average Information algorithm - it finds which var.comps are pushed to zero constantly - recalculates variance components removing such components - fix those values and calculates the most likely value for the problematic var.comp + now 'mmer2' can handle missing data in explanatory variables as lmer + now summary of 'mmer2' has names in the variance components + A.mat, D.mat and E.mat supported for polyploids + mmer can run GWAS for polyploid organisms - the models implemented are the same than Rosyara (2016): - "additive","1-dom-alt","1-dom-ref","2-dom-alt","2-dom-ref" + eigen decomposition to accelarate genomic prediction based on Lee (2015) has been added in the argument 'MTG2' of the AI, mmer and mmer2 algorithm ######################## version 1.3 (2016-03-09) ######################## + The 'bag' function for bagging-GBLUP from Abdollahi-Arpanahi et al. (2015) has been added: - The function takes a model fitted and creates a bag matrix with the top markers (most significant) and creates a design matrix to be used as fixed effects in the GBLUP model to increase prediction accuracy. + 'bag' function has been equiped with stepwise selection to make sure that markers selected by "clustering" or "maximum" p.values methods provide at least a minimum increase in the prediction accuracy. + The Fisher Information matrix can be returned from the mmer function when the AI is used (default) but the argument 'Fishers' needs to be set to TRUE. + The bug for the AI algorithm when one var.comp and K and Z are diagonal has been fixed by changing to EMMA in this naive situation. + AI algorithm has been debuged to return the most likely variance components when the likelihood takes values around the maximum in a zig zag pattern. Just takes the value where the ML was found. When the likelihod follows a scale and dropping pattern the program will do the same. A warning message is emmitted. + GWAS modality of 'mmer' now adds the names of the markers of each score to keep track the value for each marker. ######################## version 1.4 (2016-03-24) ######################## + The AI algorithm will take 5 EM steps if after 10 iterations (AI) the likelihood drops suddenly, indicating that initial values were too far from real values causing a bad behavior of the likelihood. The EM steps aim to provide initials values for those problematic variance components. ONLY mmer2!!!!!! + The AI likelihood behaving in a zig zag pattern is detected only after 10 iterations and we opted for returning the ML estimators. + Minor bugs have been fixed (names in random terms). In addition, ordering random effects based on their degrees of freedom has been implemented to provide more stability to the AI algorithm. ######################## version 1.5 (2016-05-03) ######################## + Problem of "not-full rank X matrix" has been solved by reducing the X matrix column by column until is solved. + New alleles for deletions "-" has been added to the atcg1234 function + More examples have been added to the software, including maize Technow et al. (2015) data and rice Zhao et al. (2011) + We have opted for using the "EM" algorithm as default for the mmer2 function given the fact that most users using the mmer2 function would not use covariance structures. The mmer keeps using the AI algorithm as default method. + Implementation of the initial version of TP.prep function designed to select the best training population for genomic selection. ######################## version 1.6 (2016-05-19) ######################## + Addition of the Newton-Raphson algorithm to sommer. + Vignettes with several examples available by typping vignette("sommer"). + Bug in Principal Components within the function TP.prep has been fixed. + Example of GWAS for single cross hybrids has been added to the Technow data examples. + Bug on map.plot function for genetic maps fixed when having a column "Locus" as factor + Addition of winter bean data "FDdata" to exemplify the full diallel design calculation + mmer2 function now is fully functional as mmer but using data frames. All parameters can be manipulated as mmer except for the hdm function for half diallels. + The program will subset W if some markers are not polymorphic + beeper has beed added to the program to choose what sound to display when the program is don doing the calculations. + Now genetic maps can be added for a better display of the manhattan plots. + QQ plots are now displayed when doing GWAS for different models. ######################## version 1.7 (2016-05-23) ######################## + Now the AI algorithm checks the situation when there are random effects with n < p (less observations than parameters to estimate) to take the right values in AI when extreme variance components values are found. + Bug in eigen decomposition of AI has been fixed by providing smaller starting values for the variance components + In addition, when n < p is present the initial values of the variance components are reduced significantly to provide better starting values. + New bug in the FDR function fixed. Now GWAS with or without map displais the FDR line. ######################## version 1.8 (2016-06-06) ######################## + Minor changes, mainly in plotting defaults, warning messages, etc. + Addition of the phase.F1 function to create parental maps for F1(CP) crosses. + Addition of F1geno data accompaning phase.F1 function. + Argument MTG2 has been renamed as EIGEND. + EIGEND feature added to EMMA algorithm. + Bugs for errors when V(e) tends to zero have been fixed for all algorithms, warning messages have changed. + The function for finding biggest peaks "maxi.qtl" has been added. + Multiple responses can be fit using parallelization with the 'n.cores' argument ######################## version 1.9 (2016-07-01) ######################## + lmerHELP has been slightly modified. Now only uses lmer initial values when required and only used for non-square random effects. + AI has been modified to dimminish updates values when they scalate to quickly. + zig zag detection in AI2 has been implemented, now even when boundary constraint is applied there should not be any problem. + Now 'bag' function respects if a map is available and does the bagging based on the map. + 'manhattan' function added to the package. + 'eigenGWAS' function included based on Chen et al. (2016) in Heredity. + Bug in AI failing when only one variance compnent was present and was zero, making the program collapse when trying to recalculate has been fixed. + Bugs in 'TP.prep' function fixed. + Now 'EMMA' algorithm handles multiple random effects. + The first multivariate algorithm "EMMAM" has been added to sommer (a single random effect). + Vignettes have examples for multi-response model and parallel univariate models. + 'bag' function has changed name to 'hits' to avoid confusion of selecting top 10 hits with Bootstrap aggregation (bagging). ######################## version 2.0 (2016-08-01) ######################## + Multivariate algorithms "M-EMMA", "M-AI", "M-NR" have been implemented. + Minor bugs fixed. + Multivariate GWAS inplemented for any of the methods. + The function "mmer2" can run multivariate and parallel models as well. + Bug for parallel models fixed. + Eigen decomposition of additive relationship matrix implemented for multivariate models in "NR" and "AI". ######################## version 2.1 (2016-09-01) ######################## + Standard errors and Z ratios added to the summary of mixed models. + Addition of functionality for residual structures, only AR1, CS, and ARMA supported. The most flexible functions "MNR" and "NRR" are not able to deal with missing data so you should be careful because the program imputes with the mean. + atcg1234 has a new functonality to get a presence/absence matrix for each allele at each marker. ######################## version 2.2 (2016-10-01) ######################## + The LD.decay function has been added to sommer which takes a marker matrix and a map and calculates the LD decay. + In addition, the LD.decay function has the "unlinked" and "gamma" arguments that estimate the interchromosomal threshold for the gamma percentile to determine the real LD decay combined with the loess regression. + Bugs in GWAS models for univariate and multivariate forms when missing data exist have been fixed. ######################## version 2.3 (2016-11-01) ######################## + The nearest neighbor function 'nna' has been added to adjust for neighbouring plots based on Lado et al. (2013). + The dataset 'ExpDesigns' with several datasets to teach users how to analyze certain experimental designs relevant to plant breeding has been added. + A fatal bug in mmer2 and mmer of imputing data when missing data existed has been corrected ######################## version 2.4 (2016-12-01) ######################## + The IMP argument has been added to the function to allow the user to decide if the Y matrix should be imputed or get rid of missing values when estimating the variance components in the multivariate mixed models (only). The default is FALSE which means that the missing values are removed. + The gryphon dataset was included in the package to provide some help to the users that want tosee how pedigree data is used ######################## version 2.5 (2017-01-01) ######################## + Bugs in fdr function fixed + Q+K model univariate fixed + Q+K model multivariate enabled + EIGEND feature added to NR method ######################## version 2.6 (2017-03-01) ######################## + 'imp' argument added to the atcg1234() function to allow users to avoid imputation + Covariance structures for the residual component are not longer supported. ######################## version 2.7 (2017-05-01) ######################## + at(.), diag(.), and(.), g(.) functions added to be used in mmer2. + NR algorithm updated to deal with multiple variance components equal to zero ######################## version 2.8 (2017-06-01) ######################## + addition of the pin function to the package + issue with Year:g(id) type of arguments solved + more efficient EM algorithm; covariance matrices are inverted only once + EM fixed to don't calculate statistics at the end by direct inversion + rcov argument with options 'units' and 'at(.):units' enabled + AI MME-based added to sommer + new argument DI=TRUE/FALSE for deciding between MME-based and Direct inversion algorithms ######################## version 2.9 (2017-07-01) ######################## + good versions of blocker and fill.design implemented + bug fixed in PEV for D-AI algorithm + minor bug fixes from summary functions in parallel models + going back to zero constraint where a model is fitted again witout the random effect close to the boundary. ######################## version 3.0 (2017-09-01) ######################## + D.mat function now provides 3 different calculation methods + cleaner version of NR and AI algorithms + fixed EIGEND feature + AI algorithms now takes EM steps for better initial var.comp values + MME-base algorithms removed from mmer functions + remove of several function + move of GWAS functionality to a different function called GWAS and GWAS2 + change of examples to show the flexibility + implementation of us(trait) and diag(trait) functionalities in mmer2 ######################## version 3.1 (2017-11-01) ######################## + pin function improved to don't work with scaled parameters + dominance relationship matrices corrected + general maintenance to several algorithms for minor bugs + constrained parameters in multivariate models not included in the summary as expected + better documentation about multivariate models ######################## version 3.2 (2018-01-01) ######################## + variogram and plot.variogram functions enabled to visualize the residuals or the spatial model fitted (in case of 2D spline models) ######################## version 3.3 (2018-03-01) ######################## + spl2D function to fit 2 dimensional spline model has been added and documented + spatPlots function added to visualize the fitted values and residuals + now the mmer and mmer2 function can use weights ######################## version 3.4 (2018-06-01) ######################## + and() function has been replaced by the overlay() function effectively. + spl2D() can be used directly in the formula solver mmer2(). ######################## version 3.5 (2018-07-01) ######################## + h2.fun returns now corrected heritabilities for Oakey (2006) and Cullis (2006) formulas. ######################## version 3.6 (2018-09-01) ######################## + some documentation changes + almost finished c++ version, just wait for it. ######################## version 3.7 (2018-09-01) ######################## + C++ (Armadillo library) implementation done. + now the vs() function is the main function to create variance models + ds(),us(),cs(),at() for complex variance models added + constraint matrices can be easily indicated with unsm(), fixm(), uncm() in the Gtc argument of vs() + fatal bug in the multivariate BLUPs has been fixed + implementation of predict() function for lsmeans + multivariate GWAS implemented again, fixed and ready to rock + random regression implemented + anova() function enabled for single model to get sum of squares and MS + EIGEN functionality not available anymore, same with EMMA called from mmer + mmer2 and GWAS2 have been deprecated, now mmer and GWAS functions can do all ######################## version 3.8 (2019-01-01) ######################## + stable version of c++ implementation, all bugs reported by users in 3.7 have been fixed + implement overlayed and leg in fixed effects ######################## version 3.9 (2019-04-01) ######################## + documentation on the changes from old to new versions of sommer added to vignettes among better structure of the documentation + atcg1234() function now can take a marker matrix and a matrix with reference alleles and do the conversion with the customized reference alleles. + bug in the fixed effects depending of the order has been fixed. Is due to a strange behavior of the model.matrix() function. + Now Gtc and Gt arguments can take a list of matrices to apply different constraints when using the ds() and us() functions, although the cs() can provide the same results. + Now unsm(), uncm(), fixm(), fcm() have the rep argument to repeat the constraint matrix multiple times + Now DF for ######################## version 4.0 (2019-07-01) ######################## + No updates for now other than documentation + Bug in the GWAS function for P3D=TRUE/FALSE fixed ######################## version 4.1 (2020-06-01) ######################## + Bug for removing the intercept in models has been fixed + Better documentation for GxE models + Addition if the unsBLUP function to extract right BLUPs for unstructured models + Change of the vs() argument from Gt to Gti to specify initial values ######################## version 4.1.1 (2020-10-01) ######################## + Redesign of the predict function to predict full grids + Better structure of the documentation + Bug in predict for models that include spl2D() or overlay() functions fixed ######################## version 4.1.2 (2021-02-01) ######################## + Addition of the fitted() function to return Xb + Zu.1 + ... + Zu.n instead of just Xb + Bugs reported for the predict() function addressed + The function residuals() now returns e = y - Xb - Zu instead of e = y - Xb + A.mat, D.mat and E.mat functions are now C++ Armadillo implementations + Addition of the H.mat function in C++ Armadillo + When providing and Gu argument in the vs() function, missing levels in the data vector are added to provide all BLUPs + The predict function now provides the correct standard errors + GWAS C++ implementation + Progress bar for GWAS function added ######################## version 4.1.3 (2021-04-01) ######################## + Addition of the gvs() function to fit indirect genetic effects and other competition models + Addition of an example of indirect genetic effects to vignettes. + Bug in predict() when missing data in covariates (x) exist was fixed. ######################## version 4.1.4 (2021-07-01) ######################## + Addition of the stepweight argument to mmer() to control the magnitude of the update of the AI and NR when the information matrix goes out of boundaries too quickly. + Addition of emupdate argument to mmer() to control the use of expectation-maximization updates as alternative to the use of second derivative methods (NR, and AI). + Addition of the percChange output to the mmer() function to check the % change in the variance comp. + Argument buildGu for the vs() function added to avoid rrBLUP models to become slow. This implied modifying the MNR.cpp function to understand when the matrix multiplications with K (covariance matrix) should be avoided. ######################## version 4.1.5 (2021-11-01) ######################## + Addition of the spl2Db and spl2Dmats() functions to fit all models capable to fit by the SpATS package. + Rename of spl2D function to spl2Da plus change of names for the arguments. + Fixes to a*b fixed model issues and predict improvements thanks to Sam Rogers and Julian Taylor. + GWAS by GBLUP examples added to the QG vignette. In addition new vignette for spatial modeling. ######################## version 4.1.7 (2022-07-01) ######################## + Addition of the W argument to allows users to input a weights matrix instead of only a vector. + Henderson mixed model equation version of the AI algorithm based on Jensen, Madsen and Thompson (1997) finally availble with the mmec function. Please try it. ######################## version 4.2.0.1 (2023-01-01) ######################## + P3D argument enabled again + Bug in mmec() for using weights with simple residual structures has been fixed. Now stage-wise analysis should work well. The problem still exist in the mmer() function though, where variance components are restricted in a different way. ######################## version 4.2.1 (2023-04-01) ######################## + Predict function for mmec() now works well and returns the same results than asreml. + The predict function for mmer() on the other hand has been modified to calculate C = t(W) Vi W and Ci = solve(C), where W=[X Z] but when Vi is singular it makes difficult to come up with standard error identical to the mme-based methods. + The overlay() function now works in the fixed part of the formula. + The predict function now works with models using the overlay function. + A bug in the atcg1234 function that was causing failure when few markers were used has been fixed. + Weights can be used now in the GWAS function. + The emWeight argument now uses a decreasing logarithmic scale to assign weights to each iteration of the AI algorithm to guarantee convergence as best as we can. + Bug in the spl2Dc function in mmec has been fixed ######################## version 4.3.0 (2023-05-11) ######################## + Bug in the rrc() covariance structure function when using a relationship matrix has been solved. + The new function redmm() function to create a reduced model matrix to fit huge models has been added ######################## version 4.3.1 (2023-06-11) ######################## + There was a bug where the A matrix was not being reordered when provided in the mmec() function. It has been fixed. + A function to calculate effective population size based on allele coverage has been added. ######################## version 4.3.2 (2023-08-01) ######################## + Bug in isc function fixed when matrices have a single column. + redmm() function improved calling the RSpectra package + now the mmec() function can use the different covariance structures in the fixed effect part so overlay and at models can be fitted as fixed. + Bug in atc() function fixed when a single level was used. + The covc() function has been added to calculate covariance between random effects as long as the incidence matrices of the random effects have the same levels. ######################## version 4.3.4 (2023-04-01) ######################## + thetaC and theta arguments in special functions in mmec are better documented and have some examples + the atcg1234backTransform function has been added to bring back from numbers to letters. ######################## version 4.3.5 (2024-08-01) ######################## + we have added the tps function to produce tensor product splines from the TPSbits package in a more straightforward way to produce incidence matrices that can be used in the mmec solver more easily. + keep improving the documentation of parameters and examples in different functions + rrc function for reduced rank models changed the specification and a good example is provided. + mmec function now ensures that data is transformed to as.data.frame() to avoid issues with tibble objects as reported in github. ######################## version 4.3.7 (2025-02-01) ######################## + enhancing documentation ######################## version 4.4.1 (2025-04-04) ######################## + unified mme and di algorithms unified in mmes() + removed the first two vignettes to make sommer lighter. ######################## version 4.4.2 (2025-06-04) ######################## + now the direct inversion algorithm takes as input the sparse matrices to avoid overloading R memory with dense matrices. We transform them internally to dense matrices to make operations. + bug in getPEV fixed. Now the argument works again. ######################## version 4.4.3 (2025-07-29) ######################## + updated version with better documentation for multi-trait models ######################## version 4.4.4 (2025-12-31) ######################## + MNR function exported and documented for users wanting to specify their matrices directly ######################## version 4.4.81 (2026-09-30) ######################## + ai_mme_sp2 function created to solve computation issues in the henderson algorithm. + both function mmes and mmer are back and trying to split the direct inversion from henderson once again + now we can fit many different covariance structures using mmes + more vignettes to explain covariance structures and theory behind + D.mat(nishio=TRUE) now implements Nishio and Satoh (2014), using the orthogonal dominance coding of Vitezica et al. (2013): -2q^2, 2pq and -2p^2, scaled by sum((2pq)^2). The earlier code centered heterozygosity by its observed column mean and divided by sum((2pq)^2), so D matrices (and E.mat with A#D or D#D) will differ from those of earlier versions. nishio=FALSE (Su et al. 2012) is unchanged. + D.mat and E.mat help pages corrected to describe the formulas that are used. + GLMM (PQL): any family with a valid link, including links with negative mu.eta (e.g. Gamma inverse). New binm(successes, failures/trials) builds a single proportion response whose trials are used as binomial prior weights. Pearson dispersion reported for free-dispersion families; negative binomial theta estimated with pqlControl=list(estimateTheta=TRUE). Working-model llik/AIC/BIC are NA and likelihood ratio tests are refused for PQL fits; residuals() gains type="pearson". + BUG FIX: theta_se of the Henderson engine was the inverse of the EM-blended update matrix on the -2logL scale (about half the true variance). It is now 2*inverse(AI) over free parameters, matching the direct engine; standard errors in covParNative/covparams_mmes_se change accordingly. + New wald_mmes(): incremental and conditional (marginality-respecting) Wald tests with denominator df "none", "residual", "satterthwaite" or "kr" (Kenward-Roger). anova(object) with a single model returns it. + predict.mmes: new levels argument stored in a Dtable list column, sed/avsed, pairwise comparisons (all pairs or against a reference level) with p-value adjustment and residual/Satterthwaite/KR df for fixed-effect contrasts. + BUG FIX predict.mmes: "include and average" divided each cell by the total number of cells of the term instead of the cells present for the prediction row (e.g. 1/12 instead of 1/3 for V:N); results now match emmeans. Prediction labels no longer lose characters matching the classify name. + mmes objects store call and inputArgs so model inputs can be rebuilt. + New metricm(): exponential, power, gaussian, spherical and circular distance-based correlations, isotropic or anisotropic (product per axis or geometric angle/ratio), with analytic derivatives (ASReml exp/iexp/aexp/ gau/igau/agau/isp/cir). maternm() gains geometric anisotropy. + New strm(): covariance structure spanning several random terms (ASReml str), sigma2 * K_terms x K_inner x A, with any constructor over the terms and levels aligned on the union or on Gu. covm() is its two-term case. + New vcc argument of mmes(): equality and scaling constraints between covariance parameters (ASReml vcc), identified through the new vcParams table (returnParam=TRUE and fitted objects). Implemented in both engines by projecting each AI update on w = o + T phi; theta_se follows the constraints. + New meffects_mmes(): marker effects recovered from a GBLUP term (ASReml mef) reproducing A.mat() or a custom transform, with optional standard errors and back-solved GWAS tests. + Wald, residual df and PQL dispersion use sparse rank computations. + New stackTraits() converts wide trait columns to long format with an explicit record key. mmes() accepts that key in the residual structure to pair trait residuals without relying on row order. + New familym() supports mixed-family multivariate PQL models, including binomial/Poisson/negative-binomial traits with fixed residual dispersion and Gaussian traits with estimated residual variances. corgm() gains a heterogeneous-variance mode for residual trait covariance structures. + New covmatrix_mmes() reports term covariance and correlation matrices with delta-method standard errors. Default usm()/dsm() starting shapes use per-coordinate residual mean squares. + BUG FIX Henderson compact residual score used an incorrect zero log-determinant derivative, which could prevent complete multi-trait usm fits from converging. Weighted Kronecker residual derivatives now use the same inverse-Cholesky weighting as the effective residual covariance. + BUG FIX weighted Henderson fits with non-diagonal Kronecker residuals now match the direct engine for covariance parameter estimates. + New mmes(solveOnly=TRUE, covPar=) solves the mixed-model equations at known covariance parameters (genetic evaluation) with a matrix-free, block-Jacobi preconditioned conjugate gradient that never assembles C. covPar takes a fitted mmes object, a list of natural parameters or G0/R0 matrices, or a parameter table. 1M animals (pedigree A-inverse): ~10 s single trait, ~40 s three traits. + vsm() no longer builds a dense records x levels mask per covariance coordinate, and mmes() builds the fixed design from distinct covariate rows and pairs residual records without ave(); input preparation for large data is several times faster and uses far less memory. ######################## version 4.5.13 (in development) ######################## + Added post-fit nominal unsigned deviance residuals for model checking, using stable conditional log probabilities without clipping tail losses. Existing category-wise response/Pearson residuals are unchanged; working and quantile residuals remain explicitly unsupported. These descriptive diagnostics do not provide calibrated tests or require additional fits. + Added bounded nominal predict(re.form="population") category means using deterministic Gaussian quadrature for shared/correlated category-specific unit random intercepts with identity relationships, including new groups. Up to four categories are supported; memory and 24/32-rule agreement are checked explicitly. Population SEs and averaged logits remain unsupported. Independent integration checks do not establish statistical calibration. + Added opt-in nominal predict(se.fit=TRUE, uncertainty="full") using marginal parameter covariance, covariance-coordinate mode movement and conditional random-mode posterior covariance. Existing default prediction SEs are unchanged. Computations are post-fit with curvature/memory guards; deterministic delta-method checks do not establish statistical coverage. + Stabilized nominal unstructured covariance precision scores using triangular-factor derivatives and exact factor log-determinant derivatives, avoiding cancellation near singularity. The seed-840001 regression checks a small final score and agreement with independent finite differences. At its ill-conditioned optimum, platform roundoff can produce nlminb's "false convergence (8)"; the regression verifies the warning and diagnostic status in that case without relabeling the fit as converged. Added nominal optimizer/score, covariance-conditioning, saturation and uncertainty-status diagnostics with actionable summary messages. No new dependencies, fitting-time Hessians/PEV or Gaussian/scalar-route changes. + Fixed nominal preparation tests to leave optional runtime BLAS thread control disabled, so they also run without a supported BLAS controller. + Imported stats::residuals for mixed-family Laplace residual dispatch, resolving the undefined-function R CMD check NOTE. + Started the ordered categorical milestone with internal fixed-cutpoint cumulative-logit/probit likelihoods and analytic predictor derivatives through fourth order. Added binary-limit, probability, tail, narrow- interval and full marginal-gradient checks. Public ordinal fitting remains unavailable pending public fit integration and category-aware outputs; the staged constructor now errors on placeholder GLM calculations. No Gaussian solver paths, dependencies or fit-time uncertainty changed. + Added internal joint ordinal cutpoint estimation with first-location and log-positive-gap coordinates, analytic full Laplace scores, once-per-context fixed-design location-identification checks and absent-category guards. Public ordinal fitting remains disabled until its downstream interfaces and post-fit uncertainty are connected. + Added internal single-family ordinal response/control preparation using the shared observation map and covariance-design preparation. Preserves ordered category labels, validates contiguous integer codes, initializes cutpoints from retained counts and removes only the intercept design column when cutpoints are estimated. Fixed cutpoints keep their intercept and can retain absent ordered-factor levels. Public fitting remains gated. + Fixed native GLMM mode-Hessian row scaling when a curvature weight is exactly zero. Sparse diagonal multiplication replaces in-place sparse iterator mutation, which could invalidate the iterator and raise an out-of-bounds exception. Tested on saturated ordinal and binomial rows. + Enabled complete experimental single-family ordinal logit/probit native Laplace fits: labelled fitted/predicted category probabilities, explicit latent/expected-code outputs, ordered-factor simulation, category-aware residuals, likelihood comparisons and bootstrap refits. Added post-fit natural-cutpoint SEs and full joint-parameter/mode prediction uncertainty and PEV, with no numerical Hessian or PEV during fitting. Population probabilities use normal convolution (probit) or existing Gaussian quadrature (logit). Public-interface tests include independent likelihood, integration and delta-method references. Ordinal PQL and familym remain unsupported; broader statistical coverage calibration is still pending. + Completed external targeted ordinal calibration: 3,000 paired datasets, 6,900 total fits across both links and regular/sparse/short-cluster/ weak-variance designs. All 1,200 regular fits converged and recovery gates passed, but 18/712 registered gates failed, principally probability coverage/RMSE; the full inference envelope is not qualified. Sparse absent-category rejection and weak-variance uncertainty failures remain explicit. Independent scalar-Laplace checks localized one comparison failure to external reference precision, without relaxing its gate. Package fitting, dependencies and Gaussian solver policies are unchanged. + Completed a fresh 1,200-dataset ordinal probability-interval study: Wald/logit/probit/arcsine delta methods failed 8/3/4/5 of 96 registered pointwise cell gates respectively. No method qualified across the full envelope; existing SEs and public API remain unchanged. Genuine refit-bootstrap interval calibration is still pending. + Started the nominal multinomial expansion with an internal zero-reference softmax likelihood, joint observation-block score/curvature, and directional third/fourth-order derivative contractions. It is not a set of independent binomial rows; public nominal fitting remains gated pending block Laplace mode/score integration and complete downstream interfaces. No dependencies or Gaussian/scalar-family fitting paths changed. + Integrated internal multinomial conditional modes with full category-block random-effect Hessians, fixed supplied random precision and normalized-prior Laplace likelihoods. Persistent CHOLMOD symbolic reuse retains structural zeros through saturation; input/workspace guards and explicit numerical failure diagnostics are included. Marginal parameter scores and public nominal fit/output assembly remain gated. No fitting-time uncertainty, additional package dependencies or Gaussian solver-policy changes. + Added analytic internal multinomial marginal Laplace scores for fixed predictor and random-precision parameters, including mode movement and both prior/Hessian log determinants. Sparse adjoint and bounded panel solves avoid full inverses, mode Jacobians and fitting-time uncertainty. Proposed precision matrices reuse symbolic factors when possible; invalid derivatives, insufficient workspace and numerical failures remain explicit. Public nominal response/design preparation and complete fit/output assembly remain gated; dependencies and existing solver routes are unchanged. + Added internal nominal response/reference preparation with preserved labels, category-major row maps, category-specific fixed coefficients and shared/category-specific random designs via existing covariance formulas. The nominal objective now consumes existing covariance descriptors and optimizer-coordinate derivatives, including fixed flags and relationship precision, through the current modular controls. Repeated objective/score calls reuse successful modes; derivative matrices remain term-local. Public fitting and downstream output assembly remain gated. The theory guide explains reference categories, softmax and covariance choices in beginner-friendly terms. No dependencies, fitting-time uncertainty or Gaussian/scalar solver-policy changes. + Completed public nominal multinomial Laplace fit assembly with reference- coded category probabilities, category-specific fixed effects, shared or category-specific random effects, likelihood/AIC/BIC, and nested fixed- effect comparisons. Added conditional/zero-random prediction, probability- scale prediction standard errors, factor-preserving conditional/population simulation, response/Pearson residuals, summary/vcov, post-fit PEV and nominal-specific population parametric bootstrap refits. Population- marginal prediction and deviance/working/quantile residuals remain explicitly unsupported. Uncertainty is post-fit only; no package dependency or Gaussian/scalar solver path changed. + New boot_mmes() performs post-fit population parametric bootstrap for native Laplace fits, including nested count-fixed-effect comparisons. Refits use native first derivatives without Hessians or PEV. Failed replicates, boundary flags and warnings remain explicit; incomplete tests return failure-sensitivity p-value bounds rather than silently dropping failures. Percentile intervals are descriptive on optimizer coordinates. Population simulation no longer populates the fitted relationship matrices' factor caches by reference. + BUG FIX explicitly import stats::simulate so the Laplace simulation method registers when the namespace is loaded without stats attached. Exclude configure-generated src/Makevars from source archives on all platforms; Unix configuration and Windows Makevars.win remain unchanged. + Added an opt-in zero-inflated Poisson family for native Laplace ML, with an estimated or fixed intercept-only structural-zero probability. The Poisson and zero components share no predictor coefficients. zipoisson(zero=~covariates) adds a separate fixed-effects logit predictor, including offsets, trait-aware missingness, new-data coding, simulation, residuals and joint post-fit uncertainty. Default zero=~1 preserves the scalar interface. Zero random effects, zero-inflated NB2, hurdle models and PQL support remain out of scope. + Expanded native Laplace families to grouped binomial logit/probit/cloglog, Poisson log, native nbinom2() log, Gaussian identity and Gamma log. Long-format familym() combines them through shared random effects with conditionally independent observations; an optional trials-column interface handles mixed grouped responses. NB2 shape and Gaussian/Gamma dispersion are estimated per trait by default, with starting/fixed values in task$glmm$familyParameters. Full native predictor/nuisance derivatives, fitted-family residuals/prediction, parameter counting and valid-curvature delta-method uncertainty are connected. Native Laplace inference now exposes guarded fixed-effect covariance, natural-scale covariance-parameter standard errors, Wald tests and restricted nested fixed-effect likelihood comparisons. Variance-component LRTs are refused. Conditional prediction for known random-effect levels, zero-random-effect prediction, and a restricted random-intercept population mean are available. Family-aware conditional response simulation, population random-effect simulation with a dense-memory guard, and randomized quantile residuals are available; simulation does not include parameter-estimation uncertainty. Existing PQL and both Gaussian C++ objectives remain unchanged. + Added an opt-in native ML Laplace GLMM likelihood/mode module. It consumes sommer-prepared designs and CovarianceFactor random covariance/relationship precision structures; the Gaussian ai_mme_sp2() and ai_reml_direct_sp2() objectives are unchanged. Initial family support is binomial(logit) and Poisson(log), with structured residual covariance and grouped-binomial trials not yet supported. PQL convergence is recorded explicitly, including convergence on the final allowed outer iteration. + Modularized native Laplace fitting with actual-family preparation-only setup, reusable family/precision modules, invariant relationship and descriptor caches, and reusable CHOLMOD symbolic analysis. nlminb uses native full Laplace first scores including mode/determinant dependence; the uncertainty Hessian remains numerical. execution$denseMemoryMB now controls bounded dense derivative solve panels on the Laplace route. Panels avoid complete inverse matrices; sparse Kronecker blocks are still assembled. Fit diagnostics expose scores and native work counters. + Hardened native Laplace preparation and inference: fitted contrasts, transformed prediction terms, factor levels and missing-data policies are preserved. Native count/precision validation and mode failure reasons are explicit. Removed blanket covariance-coordinate bounds and path-dependent outer mode caching. Invalid curvature, non-convergence and near-boundary fits no longer yield success-shaped standard errors. Unused Gaussian controls and unsupported prediction requests are rejected; working residuals now use the conditional link derivative correctly. + Deferred all native Laplace optimizer-Hessian uncertainty and coefficient PEV calculations until post-fit requests. Laplace now requires output$inverse=0; postPEV(), vcov(), summary() and covparams_mmes_se() perform their work after fitting. Added conditional, zero-random-effect and restricted population-mean prediction standard errors/full covariance, with memory guards from output$panelSize and execution$denseMemoryMB. Population random-effect simulation now samples through sparse relationship-precision factors, and defaults its memory budget to execution$denseMemoryMB. Exact Gaussian marginal-GLS uncertainty and seeded known-covariance interval/predictive calibration regressions cover the post-fit methods. No package dependencies or Gaussian objective changes were introduced. ######################## version 4.4.90 (2026-10-13) ######################## Engine speed work for ai_mme_sp2 (details and measurements in src/ai-mme-sp2-acceleration-plan.md). + New mmes_control(experimental=list(profile="force")) records per-phase seconds and workload counters in fit$engineDiagnostics$profile. New profile_mme_workload() tabulates them; run_mme_benchmark() carries phase_* columns; benchmarks/profile_phase0.R profiles AR(p), SAR/CAR and dense-Gu fits. + New mmes_control(memoryPolicy="speed"|"limit"), default "speed": when the CHOLMOD selected-inverse caches do not fit denseMemoryMB the fit stops and reports the memory needed, instead of a silently slower fallback ("limit" keeps the old fallbacks). + The block-banded engine is no longer chosen automatically (up to 80x slower than generic CHOLMOD on AR1 fits); it remains available with experimental=list(blockBanded="force"). + General dense block-Schur admission now compares predicted sparse-factor work (AMD symbolic fill) with dense-plan cost instead of block density. Coupled plans require a work/cost ratio of at least 0.05; uncoupled groups retain their solve/trace fast path. Invalid estimates do not admit Schur. An isolated low-fill random-graph fixture improved from 2.968 s to 0.641 s, with maximum relative output differences below 1.6e-9 across nine fixtures. Chain/latent/banded policies are unchanged by this guard. The CHOLMOD dense inverse-block cache still skips sparse groups. The provisional 64/1000 fragmentation rule was removed after true on/off validation contradicted the earlier memory-budget proxy comparison. + The positive-definiteness check of single-factor covariance structures with >= 128 effects (SAR, CAR, large AR/Toeplitz) uses a shifted Cholesky test instead of an eigendecomposition; identical fits, 2-4x less time in the constraint step. + mmes() starting values for variance components no longer use a dense QR of the fixed-effects design (a 12000 x 1000 design spent about 90% of the wall time there with multithreaded BLAS); a sparse normal-equations solve is used for well-conditioned full-rank designs. Same converged estimates, about 6x faster end to end on that model. + New mmes_control(experimental=list(blockSchur="auto"|"off"|"force")) to disable (off) or bypass the density/group-size rules of (force) the block elimination engines; engineDiagnostics$profile gains traceFallbackSolves. + Opt-in experimental=list(traceBatch="force") batches exact LDLT random-score fallback solves in independent OpenMP panels, without the shared solve lock. A 1000-level AR2 fixture improves from 7.959 s to 6.084 s with four threads; likelihood, scores, AI, estimates and uncertainty match the old exact path. Peak RSS increases about 14.2 MiB. One-thread fits do not benefit, so the default stays off. Panel scratch obeys speed/limit memory policies. New profile counters: traceBatchSolves and traceBatchWorkspaceBytes. + Opt-in experimental=list(sparseFactors="force") uses exact sparse AR(p)/ANTE precision operands in derivative products and tensor applications, compact ANTE innovation rows, and sparse SAR/CAR covariance/derivative solves. On paired 1000-level fixtures, whole-fit medians improve from 6.287 s to 4.770 s for SAR and 5.699 s to 4.168 s for CAR; peak RSS falls about 114.9 MiB and 45.0 MiB, respectively. Likelihood, scores, AI, estimates and uncertainty agree within 5.23e-9 relative error across ten fixtures. AR/ANTE whole-fit gains are modest or unestablished, so defaults stay off. Dense compatibility and covariance-derivative storage remain; no derivative thresholding, PCG estimator change, or optimizer change is introduced. + LDLT selected inverse indexes dense columns by position instead of hashing (solver="ldlt" with a dense Gu, 300 individuals: 72 s to 16.5 s). + ai_mme_sp2 now delegates covariance evaluation, precision, derivatives and tensor operations to a shared covariance engine. Structured kernels use representation capabilities rather than model labels, including custom modified-Cholesky innovation callbacks. Saved built-in factors acquire missing capabilities at the R compilation boundary. Path and workspace policy remains in mme_paths.h; existing acceleration defaults and output names are retained. + BUG FIX complete-block ANTE residuals with sparseFactors="force" no longer reject valid covariance derivative indices. Multi-factor SAR/CAR scores now include the Kronecker log-determinant multiplicity. Sparse spatial evaluation caches obey their aggregate entry allowance without unsigned underflow. + rrm() now fits genuine reduced-rank covariance with all specific variances fixed at zero, matching the covariance family of ASReml rr(). The first loading is fixed to one to identify the overall scale supplied by vsm(). Native latent-factor Henderson and direct-inversion fitting support the singular random covariance without an artificial diagonal remainder. + fam() gains variance="homogeneous" to estimate a common specific variance (the former rrm() covariance family); the default remains "heterogeneous". loadings_mmes() returns exactly zero specifics for RR and scales specific variances by sigma2, not sqrt(sigma2). RR regression scores recover latent BLUP coordinates; Bartlett scoring rejects zero-specific RR models. + Native latent RR fitting now uses loading-aware scores and average information instead of BFGS and numerical observed information. Covariance derivatives include changes in the latent design; the original covariance-family descriptors and sparse Henderson solves are reused. Parameter uncertainty uses final unblended AI, while stepWeight and emWeight control updates and the iteration monitor records accepted fits. + Native RR AI traces now share exact design and residual contractions across covariance parameters, avoiding repeated observation-space derivative and relationship solves. Conditional likelihood trials skip unused ordinary trace preparation. Verbose RR fits report their MME path, coefficient solver (including eligible block-Schur dispatch), and accepted iteration progress. + RR progress prints the coefficient engine before the table header and includes wall/CPU times, EMweight and actual accepted trust/line-search scales. Numeric values are retained in engineDiagnostics$iterationProgress; inapplicable boundary-restraint and notPD columns remain omitted. + BUG FIX native RR AI fitting now supports dsumm() spatial residuals. Covariances, analytic derivatives, and shared trace contractions operate on observed residual blocks rather than expanding the section-by-grid covariance. Existing section slicing and factor derivatives are reused. + Native RR AI now reuses the live LDLT/CHOLMOD or block-Schur coefficient solver and the likelihood's residual precision operator. Spatial fits use the shared grid engine and its factor-wise derivatives rather than a second dense observed-block precision. Random traces cancel relationship covariance actions through the latent Henderson identity. Verbose output identifies the shared residual engine; AI phase timings are retained. + Opt-in mmes_control(acceleration="hybrid-ai") reduces repeated Henderson sensitivity solves using Powell-damped BFGS between exact AI refreshes. Scores, likelihood acceptance, bounds and PD safeguards remain unchanged; intermediate stabilization uses the BFGS diagonal. Final InfMat and covariance-parameter uncertainty always use exact AI, including at the iteration limit. Convergence requires exact step and projected-score checks. Gaussian LDLT/CHOLMOD fits are supported; vcc ties explicitly retain exact AI. Other optimizers and the default behavior are unchanged. Diagnostics expose refreshes, resets, certifications and sensitivity RHS counts. + Experimental mmes_control(experimental=list(saddlePoint="force")) adds an exact augmented Henderson MME solver using pivoted Eigen sparse LU. Defaults and "auto" remain unchanged. Backend-independent exact panel contractions and normal-equation output assembly are reusable in mme_exact.h; the augmented factorization lives in mme_saddle.h. REML/ML retain exact scores, AI, covariance SE and coefficient uncertainty. Fitting avoids normal-equation cross-products; final C is assembled once for prediction and saved-object compatibility. Unsupported routes fail explicitly. Initial benchmarks do not establish a speed or memory gain; the solver remains force-only, with no automatic admission rule. + mmes() now blends genuine EM/GEM and AI proposal directions rather than information matrices. blendWeight replaces emWeight (a deprecated named alias), and blend="em" selects this method. Scalar, fully free diagonal and unstructured terms use closed-form posterior-moment updates when available; other structures and constraints use numerical increasing-Q generalized EM. Complete unstructured residual blocks also have closed-form M-steps. Mixtures use likelihood backtracking with an EM fallback. Full EM weights skip iterative AI, zero weights skip EM, and final inference always uses exact AI. Positive weights require deterministic Gaussian Henderson REML or ML without other accelerators, factor-score augmentation, PQL or saddle-point fitting. Unsupported modes default to pure AI when blendWeight is omitted; explicitly positive weights remain errors. Blended fits default to stepWeight=1; pure AI retains its previous damping. engineDiagnostics$blend reports work counts, Q gains and numerical M-step convergence. Exact moments obey the denseMemoryMB budget. Henderson weight factorizations now consistently preserve observation order across EM and AI. This fixes a previous correlated-weight ordering discrepancy; affected weighted estimates can differ from earlier versions. A scalar whole-number blendWeight greater than 1 now sets the exponential decay length from 1 to 0.03, capped at nIters, followed by constant 0.03. Fractional counts and vectors containing values above 1 are rejected. + blend="pxem" now selects scalar parameter-expanded generalized EM (PX-GEM) blended with AI. Each random term has a positive expansion scalar; joint numerical optimization preserves all covariance families, fixed values, scale floors and vcc constraints in original reduced coordinates. Existing EM/GEM updates seed the expanded M-step. Posterior moments include fixed/random and cross-random uncertainty, and reuse existing MME paths. PX proposals use observed-likelihood backtracking with ordinary EM/GEM fallback. Full weight skips iterative AI; zero weight skips EM/PX work. Final inference remains exact AI. Diagnostics expose expansion factors, additional Q gains, numerical convergence/evaluation counts and fallbacks. Additional moment/optimizer workspace obeys denseMemoryMB. This is scalar numerical PX-GEM, not full matrix expansion; fewer iterations do not guarantee lower runtime. The default remains blend="em". Reduced-rank EM/PX-GEM now cache exact loading-basis posterior moments once per E-step instead of repeating coefficient solves at every numerical M-step objective probe. This fixes severe first-iteration latency in models with many loading and residual parameters, without changing Q. EM/PX random moments now reuse exact Schur/inverse-block and selected-inverse contractions, requesting only diagonal moments for diagonal covariance updates. Scalar residual moments reuse a fixed-effect-aware trace identity. Scalar non-reduced-rank PX expansion moments reuse sparse coefficient contractions when memory permits. Generic bounded exact panel solves remain the fallback. New work counters expose contraction and fallback usage. AI directions inside EM/PX schedules now use native exact AI kernels for non-reduced-rank models instead of generic latent information calculations. Zero-weight iterations skip moment callbacks. Mixed iterations share native AI factorizations and inverse caches with posterior-moment work. Accepted exact-information trials supply subsequent zero-weight iterations and final inference without duplicate factorization. Reduced-rank AI retains its shared latent kernels; blend guards and constraints remain unchanged. + Omitted blendWeight now selects AI-first optimization with on-demand EM/GEM recovery (or scalar PX-GEM with blend="pxem") for eligible Gaussian Henderson REML or ML fits. Healthy AI iterations skip EM/PX moment and objective work. Repeated poor progress, failed AI line search, or unreliable information solves trigger one recovery step when the projected score remains non-small. Recovery must improve likelihood before AI resumes. Automatic convergence requires a scaled projected-score certificate. Failed recovery warns and returns the last accepted point without convergence. Explicit numeric blend schedules and unsupported-engine defaults retain their behavior. Recovery diagnostics record triggers and stop reasons; automatic refits retain blendWeight=NULL. + ML now supports explicit EM/PX-GEM schedules and default AI-first recovery. ML posterior moments use the random-only D inverse, with deterministic current GLS fixed effects during covariance and loading Q/PX probes. Likelihood trials reprofile fixed effects and accept against profile ML. Random, residual, weighted and reduced-rank moment paths share a separate posterior action; joint solves and final coefficient uncertainty are unchanged. CHOLMOD retains independent C/D symbolic caches across conditional fits. Diagnostics identify ML versus REML. Dense conditional-moment and expanded-Q tests cover both likelihood modes. + mmes() iteration output now shares one compact formatter across ordinary, hybrid and reduced-rank AI, scheduled EM/PX blends, AI-first recovery and direct inversion. Columns show accepted likelihood/gain, update method, recovery policy/trigger, trust and line-search scales, notPD repairs, and elapsed seconds. Redundant wall-clock/CPU columns are removed; detailed stored diagnostics remain unchanged. Recovery rows identify the actual method tried, including late line-search recovery and EM/PX fallbacks. + Unreliable automatic AI solves now try a scaled, ridge-regularized, trust-limited AI direction before computing EM/PXEM moments. Both retries require feasible observed-likelihood improvement; ordinary AI resumes after each accepted step. Explicit numeric blend schedules are unchanged. Progress reports AI-reg and diagnostics record regularized attempts, accepted steps, source iterations, and ridges. + Standalone sparse, saddle, and PCG native tests now check for their private source headers before compiling. They skip in installed-package check layouts without src/, while source-tree runs retain the native checks. + Direct inversion (henderson=FALSE) now supports the same blend/blendWeight controls as Henderson: true observation-space EM/GEM, scalar PX-GEM, mixed directions, scalar/vector schedules, and exponential tapers. Omitted weights select lazy AI-first recovery; explicit zero preserves AI-only fitting. All direct covariance families support ML and REML, including custom factors, constrained and ragged residual structures, and zero-specific reduced-rank latent terms. Numerical M-steps increase Q; they do not claim global maximization. ML profiles fixed effects in Q, and REML includes joint fixed-effect posterior uncertainty. Full EM skips AI update solves, while final inference remains clean AI. Direct EM workspace obeys denseMemoryMB. Diagnostics record Q work, actual methods, numerical M-step status, recovery, and likelihood reuse. Henderson algorithms are unchanged; the existing numerical Q minimizer is shared without changing its implementation. + Opt-in experimental=list(reducedAI="force") now packs independent vcc directions before Henderson sensitivity solves and assembles exact reduced AI without full sensitivity or AI buffers. Default AI-first recovery and positive EM/PXEM direction blending support Gaussian deterministic REML/ML, preserving fixed/tied coordinates, boundary restrictions, trust units, regularized AI recovery and exact score certification. Full covariance reporting retains a full exact final evaluation. covarianceInformation="none" avoids report-only full information while preserving coefficient uncertainty. Default remains off. Auto admission reports unsupported-path fallbacks; force rejects them. Legacy explicit blendWeight=0, acceleration, latent reduced rank, PCG, PQL, solve-only, saddle-point and factor-score augmentation are not supported. Full score traces and active local curvature-pair work remain unchanged; small full-reporting fits can be slower. Diagnostics expose actual/skipped sensitivity columns and reduced workspace sizes. + Opt-in experimental=list(commonScale="force") profiles the overall covariance scale analytically in eligible Gaussian deterministic Henderson REML/ML fits. Independent shape sensitivities incorporate vcc ties directly; exact profile scores and AI scale Schur elimination need no extra factorization per candidate. The dedicated optimizer uses trust caps, regularized AI, likelihood backtracking and exact final score certification, not EM recovery. Product-level variance scales must be free, fixed designs full column rank, and ties common-scale compatible. Fixed shapes and positive variance-ratio ties are supported. Only interior product variances above max(1e-8,tolParInv) are supported: initial floor violations error, trial violations backtrack and stalls warn without declaring convergence. Refit with commonScale="off" for boundary fitting or ordinary AI/EM recovery. Positive blendWeight, PCG, PQL, solve-only, acceleration, latent reduced rank, saddle-point and factor-score augmentation are excluded; omitted or zero blendWeight is accepted. Default remains off. Auto admission exposes static fallback reasons rather than hiding runtime failures. A final ordinary evaluation preserves original-coordinate outputs, coefficient uncertainty and requested full covariance scale/shape information. Information omission still performs that final evaluation. Wald/KR perturbed fixed-parameter evaluations bypass scale profiling. Common-scale diagnostics expose total work and certification; no general speedup is claimed. + Breaking numerical-control API: mmes() now accepts one modular control object through control=mmes_control(). Numerical arguments and the former engine/experimental/pcg containers are replaced by task, formulation, representation, solve, logdet, score, curvature, update, proposal, convergence, output, execution and diagnostics settings. Unknown names are errors; legacy public aliases are not accepted. Model formulas, families, covariance constraints and known covariance inputs remain model arguments. See sommer.mmes.options for migration examples. Default algorithms, schedules, tolerances, work budgets and output requests retain the previous numerical bundle. Aitken extrapolation belongs to proposal; hybrid AI belongs to curvature. Their combination remains unsupported. Shared typed native policy separates those responsibilities without changing numerical kernels. Fitted objects retain control and an executionPlan; detailed native diagnostics remain the eligibility evidence. + mmes_control() and mmes() check model-independent option combinations against one compatibility table. Supplied settings that conflict, or that the selected path would ignore (for example PCG estimator settings with a deterministic solver, Henderson settings with formulation="direct", or estimation modules with task$mode="solve"), stop with one message naming every offending setting and a fix. Restated defaults are accepted. Changing task$pql on a Gaussian identity fit is an error. Fit-time eligibility messages now use module option names (e.g. solve$method, output$inverse, representation$equations) instead of removed arguments. + Consolidated legacy mmer, its methods, covariance constructors and GWAS functions in R/DEPRECATED_mmer.R without changing their definitions. Removed duplicate current-path helper definitions. + Current mmes post-fit prediction, posterior variance and marker inference reuse operation-local coefficient factorizations. Marker SE calculations form relationship-marker products by block, reuse retained precision and preserve sparse supplied inverses. + Native covariance reporting now propagates cross-term uncertainty jointly; vcDelta uses the full native covariance matrix. Unavailable uncertainty does not contaminate independent native quantities. + Reduced allocations in covariance-matrix SEs, leverage diagonals, spatial splines and APY assembly; reused covariance-factor compilation, shared control defaults and PQL warm-start compatibility checks. + Benchmark results now record the actual convergence flag, retain uniform rows on warmup/timed failures, compare final likelihoods and mark unchecked or unavailable-reference parity as NA. + Simplified the regression suite by removing overlapping family/solver smoke sweeps, detailed internal-counter and output-format checks, and unused reference helpers. Retained independent numerical references, public API validation and reported-bug regressions. Replaced spatial hash snapshots with direct tensor/design checks and reduced repeated threading stress fits. Future tests should cover a distinct public contract or numerical failure, rather than repeat the full cross-product of already tested options. + Package consolidation: separated covariance descriptor/family code and formula helpers without changing their contracts. Lifecycle hooks now live in R/zzz.R. Native utilities, relationships, post-fit inference, direct fitting and known-covariance solving have separate translation units; shared direct covariance evaluation and MME design assembly have explicit internal boundaries. + Consolidated public-to-private PCG option mappings and added ownership and mutable-control regression tests. + Added cross-platform package-check CI and a required source-tree native test job. Existing standalone native control checks now run automatically, with temporary caches and restoration of both global test options. + Reduced benchmarks to one seeded performance/parity runner and its usage guide; removed historical outputs, installed libraries and exploratory scripts. Old recorded measurements remain historical, not current files. + Added an executable README quick start and contributor/test guidance. MASS and crayon are implementation imports rather than attached dependencies; use qualified calls (e.g. MASS::mvrnorm) in user scripts. Matrix and enhancer remain attached for the established model interface. + Fixed Henderson random-effect vector u to use the original response scale, consistent with bu and uList. Added independent mixed-model equation checks across LDLT, CHOLMOD, PCG and direct fitting, and strengthened prediction fixtures to require converged fits away from zero-variance boundaries. + Gaussian REML fits now omit independent random covariance components whose designs lie entirely in the fixed-effect space. The conservative check handles free single-axis dsm/atm variances and whole free-scale terms with fixed shapes. Omitted components retain zero covariance and random effects in public outputs, with NA uncertainty, an explicit status and an audit table; numerical regularization is no longer mistaken for estimability. + Added opt-in Gaussian REML conditional FA updates through mmes_control(update=list(method="conditional-fa")). Sparse posterior moments are a separate mathematical block with per-fit symbolic reuse, refreshed numeric factors, cached sparse quadratic contractions and bounded inverse panels. Incomplete/repeated records and diagonal or FA residuals are supported; unsupported combinations fail explicitly. + Added fam(shared="group") for common genetic/residual loading axes and source-specific diagonal factor variances on the conditional path. Factorized outputs omit dense trait covariances and uncertainty matrices. This exact EM foundation is experimental, not a thousand-trait speed guarantee; convergence may be slow near variance boundaries. + Regression FA scores now use a rank-sized Woodbury solve instead of a dense trait covariance solve, including factorized conditional-fit output. + Separated latent FA equations from optimization: use representation=list(equations="conditional") with update$method="ai" (default) or "em". The earlier method="conditional-fa" remains EM. Added exact analytic REML scores and average-information contractions, joint shared-loading derivatives, reusable posterior-factor RHS solves, rank-revealing safeguarded AI steps and explicit EM recovery diagnostics. Physical-specific/allocation bounds use execution$inverseTol and report their active KKT coordinates, preventing near-deterministic latent systems from accepting numerically unreliable likelihood improvements. Conditional AI admits dense derivative/curvature workspace through the existing memory policy; ordinary AI and conditional EM defaults are unchanged. + Native Laplace ML familym() now combines ordinal, nominal and scalar response traits using observation-level categorical likelihood blocks. Mixed fitted/prediction outputs are trait-named vectors and labelled probability matrices. Random formulas remain explicit: .familyCoordinate requests category-specific effects; shared effects are not silently expanded into independent coefficients. Mixed prediction uncertainty, residuals, simulation, summary/vcov, postPEV, nested comparisons and boot_mmes are supported, with all uncertainty deferred until post-fit. Mixed nominal outer scores centrally difference the native block Laplace objective; scalar/ordinal-only scores and Gaussian/PQL routes are unchanged. + Fixed observation-space direct PQL fitting by retaining its sparse [X Z] predictor design without invoking Henderson fitting. Added direct/Henderson and independent GLM/working-model regression checks for families and links, grouped trials, offsets, filtering, structured covariance, W/weights, vcc, ML/REML, warm starts, NB shape estimation, summaries, Wald tests and PEV. Direct PQL now requires both inner and outer convergence; failed certification warns and remains visible in convergence, pqlConverged and pqlMonitor. Direct-engine post-fit support is completed in the following entry. + Enabled predict() and postPEV() for direct Gaussian and direct PQL fits. Post-fit observation-space covariance includes fixed/random and cross-term error covariance, structured and reduced-rank priors, working residual weights and public rank-reduced coefficient maps, without Henderson fitting or coefficient assembly. Diagonal PEV uses bounded panels; full covariance and prediction workspace honor the existing memory budget. Gaussian VarU and postVarU() use the same operator. Prepared sparse inputs are serializable; native factors stay operation-local. PQL inference remains a conditional link-scale working-model approximation, with VarU restricted to Gaussian fits. + Enabled random covariance vcc equalities and ratios for all native Laplace scalar, ordinal, nominal and mixed familym likelihood routes. A shared reduced-coordinate adapter expands proposals and pulls back native scores; likelihood kernels and Gaussian/PQL solver dispatch remain separate. Laplace setup now exposes its random-only vcParams without fitting. Parameter counts use independent coordinates. Post-fit covariance SE, prediction uncertainty, PEV, serialization and bootstrap refits preserve the same map; no fit-time Hessian or PEV is introduced. Family parameters and residual scales are excluded. Nested fixed-effect comparisons require identical covariance constraints. Malformed constraints, incompatible bounded correlations and inconsistent fixed anchors fail explicitly. + Enabled structured Gaussian observation-residual blocks for native Laplace ML, including Gaussian traits alongside scalar, ordinal and nominal familym responses. Gaussian subsets reuse the direct residual covariance assembler, including sectioned structures, compatible W and independent-block metadata. Residual descriptors own Gaussian scales and appear in setup and covariance reporting; vcc remains random-only. Full observation curvature and analytic residual scores feed the existing modular mode solver. All-Gaussian fits recover exact Gaussian ML. SE, prediction uncertainty and PEV remain post-fit; simulation, bootstrap, nested comparisons and residual diagnostics preserve the fitted covariance. Dense Gaussian-subset and derivative workspaces obey execution$denseMemoryMB. Discrete/Gaussian-discrete residual correlations and direct Laplace fitting remain unsupported. + Formalized correlated Gaussian latent processes for non-Gaussian families using existing random covariance descriptors and relationship precision, with temporal/spatial/shared-response recipes and explicit basis semantics. Population prediction now uses actual predictor variances or joint nominal logit covariance, supporting structured processes, slopes and nonidentity relationships instead of independent-intercept restrictions. Bounded solve panels avoid dense observation covariance. Nominal/mixed population SE and covariance propagate all estimated outer parameters strictly post-fit. Fixed a mixed nominal row/category ordering error in population integration and preserved expanded prediction relationship metadata for scalar SEs. New independent groups require identity precision; out-of-basis covariance coordinates fail explicitly. Nominal integration retains the four-category, memory and deterministic quadrature-accuracy gates. + Added interchangeable C++ random-effect and observation-space Laplace algebra, selected through task$glmm$backend (auto/random/observation). The shared mode/gradient interface preserves the same likelihood, determinant correction, covariance descriptors and full signed/block curvature; it does not call the Gaussian AI optimizer. The observation route uses sparse prior solves, bounded covariance panels and a dense predictor-space root/factor instead of assembling the random Hessian. Auto selection considers dimensions, structural coupling and memory, with explicit diagnostics and no silent numerical fallback. All scalar, ordinal, nominal and mixed paths, constraints, Gaussian residual blocks, prediction, bootstrap and saved fits preserve their statistical targets. SE/PEV remain strictly post-fit through operation-local mode/Schur solves. Added dense-reference, finite-difference and public workflow regression checks plus a reproducible one-thread benchmark script. Gaussian/PQL formulation and solver controls remain distinct from Laplace backend choice. + Added safeguarded native Laplace mode warm starts and analytic mixed nominal marginal gradients, with explicit numerical reference scores. Exact Takahashi gradient contractions and independent mode-Hessian block solves are modular opt-in random-backend paths; missing inverse entries use exact panels. ZIP mode reuse is disabled automatically; explicit reuse cold-verifies and retains the historical root. Public controls, diagnostics, post-fit inference, prediction, bootstrap and serialization preserve the choices. Warm reuse and analytic scores are defaults; panel traces/global factorization remain defaults pending workload-specific benefit. Added independent-reference tests and whole-fit CPU benchmarks. + Added opt-in curvature$assembly="augmented-working" for Gaussian deterministic Henderson AI. Prior-shifted coefficient-space working directions reuse the joint coefficient factorization, without random observation-space working variates or an explicit coefficient inverse. The existing sensitivity assembly remains the default. Structured residuals, weights, compression, reduced/tied directions, recovery, profiling and final uncertainty retain existing contracts; unsupported latent/conditional, non-Gaussian and iterative requests fail explicitly. Added independent border identities, same-point and end-to-end parity, workspace diagnostics and four-engine comparative benchmarks. + Faster reduced-rank (rrm) AI with grid residuals such as dsumm(vsm(ar1m(), ar1m(), ism(units))): residual block contractions use only the block's active coefficient columns and the selected inverse (exact solves on those columns otherwise), residual derivative traces reuse the grid derivative action, and the design transpose is formed once. Results are unchanged; a 9-environment AR1xAR1 trial fit ran about 2.6x faster. + Sparse grid residuals. Complete grid residual blocks whose factors have banded precisions (AR1, low-lag modified Cholesky) now enter C with a sparse banded pattern built from structural triplets, so the symbolic factorization is analysed once per fit (CsymbolicAnalyses = 1) and residual score traces use only selected-inverse entries inside C's pattern. Controlled by representation$sparseFactors; diagnostics gridSparseBlocks, gridSparseBlockUses and gridBandFallbacks. + New representation$gridAugment = "auto"/"off"/"force" (option sommer.mme.gridAugment). Like ASReml, missing cells of grid residual blocks (e.g. dsumm(vsm(ar1m(range), ar1m(row), ism(units)), by=trial)) are added as pseudo-observations with one dummy fixed effect each so the residual blocks are complete and sparse. Exact for REML: llik, variance components, b, u, Ci, fitted values, AIC/BIC and predictions match the unaugmented fit; dummies are removed from the returned object and engineDiagnostics$gridAugmentation reports them. EM moments exclude the pseudo-observations, the cholmod inverse cache keeps only real fixed effects in its dense border, and reduced-rank AI takes residual traces directly from selected-inverse entries (directResidualTraces) instead of dense per-block contractions. A 9-environment AR1xAR1 dsumm() trial fit went from 6.5 s to 2.1 s (cholmod) and 17.7 s to 8.2 s (ldlt); with an rrm() GxE term from 16.3 s to 7.5 s, with identical estimates. + Spatial vignette: new section on computational efficiency of spatial residuals and on the alternative AR1xAR1 random field plus nugget model. + CHOLMOD trace preparation now prefers the sparse selected inverse on augmented residual grids, avoiding redundant dense random inverse-block caches for fam() models at any supported rank. Memory admission and exact general-cache/solve fallbacks are preserved. On a 9-environment AR1xAR1 trial model, four iterations of FA2/FA3 fell from about 26 seconds to 3-4 seconds with identical likelihoods, estimates, scores and AI matrices. Added rank-1/2/3 dense-Gu FA/grid parity and inference regression tests. + A.mat() now accepts explicit ploidy with allele dosages from zero to ploidy and uses the additive scale ploidy*sum(p*(1-p)). The default retains diploid -1/0/1 coding; ploidy is not inferred from incomplete dosage classes. Mean imputation retains the original dosage scale. meffects_mmes() shares the same ploidy-aware transform for marker effects and their standard errors. + A.mat(shrink=TRUE) now computes deterministic Endelman-Jannink (2012) covariance shrinkage in C++, with the same EJ estimator as rrBLUP for matched retained and imputed markers. Supports legacy diploid coding and explicit polyploid dosages, preserves mean diagonal scaling, and reports intensity in the matrix's "shrinkage" attribute. Defaults and the existing imputed-result structure remain unchanged. Regression shrinkage and EM imputation are not included. + A.mat(method="ASV") now implements the average semivariance normalization of Feldmann et al. (2023), equation (2), in C++. Centered marker crossproducts are divided by trace/(n-1), giving unit average pairwise semivariance. Diploid and polyploid inputs retain the existing filters and mean imputation. VanRaden remains the default; ASV rejects shrink=TRUE and does not regularize the matrix. meffects_mmes(method="ASV") uses the corresponding marker transform for effect and standard-error recovery. + D.mat() and E.mat() now accept explicit ploidy and allele dosages from zero to ploidy. Native kernels use orthogonal digenic dominance under binomial dosage equilibrium, with variance normalization 2*ploidy*(ploidy-1)*sum((p*q)^2). All three Hadamard epistasis types support polyploids. Defaults and both diploid methods remain unchanged; nishio=FALSE is rejected for polyploid dominance rather than applying an unspecified Su extension. H.mat() already accepts polyploid additive relationships without a ploidy argument; its documentation now explains inheritance and scale requirements. + Experimental traditional nonlinear Gaussian mixed models now accept nonlinear=nonlinear_curve("x", "exponential"), "logistic" or "saturation". Native curve derivatives and exact Gaussian integration jointly estimate fixed mean, additive random covariance and residual variance parameters. Numerical choices use task$nonlinear in mmes_control(); point prediction, named natural-scale coefficients and serialization are supported. Initial scope is univariate Gaussian ML, Henderson, native analytic covariance descriptors and homogeneous independent residuals. Layers, backpropagation are not included. PEV and all uncertainty remain post-fit; nonlinear post-fit inference and model-comparison methods are explicitly unavailable in this slice. + Traditional nonlinear curves now support one Gaussian random curve parameter with optional independent ordinary additive effects. Positive parameters use log deviations; midpoint uses identity deviations. Dedicated exact-Hessian safeguarded modes and complete analytic Laplace scores include mode motion and third curve derivatives. task$nonlinear$mode and population controls govern deterministic starts and checked population quadrature. Conditional, zero-deviation and prior-population point means are distinct; group parameters are retained. Native covariance descriptors and known relationship precision are reused. Multiple random parameters, zero overall random variance and unseen new-data groups are explicitly unsupported. PEV/uncertainty remain post-fit. + Nonlinear Gaussian and Laplace fits now support pure reduced-rank rrm() through normalized native loading maps and proper latent Gaussian priors. Analytic loading scores include the changing design, mode and Hessian; singular full covariance is never inverted or regularized with a ridge. Full-coordinate random modes and group parameters are restored, with latentRandomMode/randomCoordinates retained separately. Population means use implied marginal prior variances, including reduced-rank covariance. Dedicated nonlinear tests cover FA, diagonal, unstructured, serial, compound-symmetry and representative Kronecker structures, with constraints, fixed parameters and known relationships. No new controls or fit-time uncertainty outputs are introduced. + Traditional nonlinear mixed models now support all existing native response families and links, including ordinal, nominal, zero-inflated Poisson zero-regression and categorical/scalar familym() mixtures. The curve is a shared link-predictor contribution, not a response-scale mean or a separate parameter set per trait/category. Modular family-aware Laplace modes use exact conditional Hessians and complete analytic fixed/curve/covariance/nuisance scores, including moving RR maps and block nominal curvature. Exact additive Gaussian ML remains unchanged. Family choices reuse task$glmm$familyParameters; other fitting choices use the existing nonlinear controls. Link/response prediction, diagnostic residuals, serialized new-data reconstruction, checked joint-prior population integration and conditional/population response simulation are supported. PEV and every uncertainty product remain separate post-fit processes; structured Gaussian residuals and multiple random curve parameters remain outside this phase. + Traditional nonlinear inference is now available strictly post-fit through post_nonlinear(), postPEV(), vcov(), confint() and prediction SE/intervals. Checked differences of analytic marginal scores provide observed parameter curvature; native conditional curvature and mode sensitivity distinguish conditional covariance from empirical Bayes approximate PEV. Natural-scale covariance, group-parameter SE and descriptor-report SE include delta-method transformations and constrained/RR coordinates. Scalar, categorical and mixed-family output shapes are preserved. Population uncertainty differentiates joint-prior integration; scalar future-response intervals include observation and prior variability. Numerical choices use task$nonlinear$inference in mmes_control(). AIC/BIC count all free coordinates; regular nested ordinary fixed terms support approximate likelihood-ratio comparisons. Explicit finite working-coordinate nulls support fixed/curve Wald tests. Boundary tests, weak curvature and unsupported linear inference formulas fail explicitly. Fitting, point summaries and generative simulation remain inference-free. + Traditional nonlinear qualification adds opt-in fitting phase timings and profile_nonlinear() for separate inclusive curve, Jacobian, conditional curvature, marginal objective/score and prediction measurements. Execution plans report actual optimizer/gradient, iteration/evaluation counts, starts, mode diagnostics, dimensions and memory admission policy. Sparse-component factor-fill bounds reject oversized Hessian operations before assembly without densifying conventional designs. Bounds remain conservative for large connected relationship graphs. A reproducible installed scaling benchmark and independent dense ML tests qualify the traditional-curve envelope; no layers or backpropagation. Validation uses numerical oracles and simulated truth without nlme or lme4 dependencies, including Suggests and CI requirements. + Traditional nonlinear curves now support multiple independent named random parameter formulas and a separate joint=list(parameters=..., random=...) covariance block. Existing random formulas and vsm() syntax are unchanged. The internal .parameter axis uses native covariance descriptors, known relationships, constraints and reduced-rank maps without replicating responses. Exact conditional curvature and complete analytic Laplace scores include every cross-parameter second/third derivative. Joint prior prediction, simulation, serialization, profiling and explicit post-fit group covariance/PEV retain the fitted coordinate basis. vcov(type="group") exposes natural-scale post-fit group-parameter covariance; uncertainty remains absent from fitting. ###################### ## PENDINGS AND PRIORITIES + Fix plot.mmes + check how dense schur engine is used when more than one random term with dense Gu is used