idiographic

Bug fixes

• glasso_kkt() no longer reports optimal zero-constrained fits as non-optimal. At an entry hard-constrained by zero, the inactive-edge condition |W_ij - S_ij| <= rho_ij does not apply: the equality constraint carries its own Lagrange multiplier, which absorbs the residual. Constrained pairs are now excluded from the check. A fit matching glasso's Fortran kernel to 2e-12 previously certified as violating optimality by 0.028.
• glasso_fit() now rejects a covariance matrix that is not positive semi-definite (classed condition idiographic_not_psd). A negative eigenvalue produced a precision matrix with a negative diagonal, from which every partial correlation was NaN. Singular but positive semi-definite covariances remain valid input. The tidy accessor also refuses a precision matrix with a non-positive diagonal (idiographic_bad_precision) rather than letting stats::cov2cor() warn and emit NaN weights.
• An asymmetric element-wise penalty matrix is now rejected by glasso_fit() and by fit_graphical_var()'s regularize_mat_kappa. The penalty on edge (i, j) is a single scalar over a symmetric Theta, so an asymmetric penalty is ill-posed rather than stricter; previously it was accepted and produced a fit that failed the package's own optimality check.
• glasso_kkt() validates penalize_diagonal instead of letting isTRUE() silently map NA and other invalid values to FALSE, and warns (idiographic_glasso_kkt_override) when a supplied rho or penalize_diagonal differs from the one the fit was made with, since the returned violation then certifies a different problem.
• The internal graphical-lasso optimality checker used to measure the *unpenalised* diagonal stationarity condition (W_ii = S_ii) even for fits made with penalize.diagonal = TRUE, whose condition is W_ii - S_ii = rho. It therefore reported a spurious violation of exactly rho for every such fit — including glasso's own Fortran output, which is how this was found. It now takes the diagonal-penalty flag into account. No estimator result and no previously published number changes; only the diagnostic was wrong.

Documentation

• matrices() now documents that it is a display verb by default, and points at print = FALSE for programmatic extraction.
• The three most computationally expensive vignettes -- Graphical VAR, Bayesian VAR/DSEM, and GIMME -- are now pkgdown articles rather than installed vignettes. Their content is unchanged and they remain published at <https://pak.dynasite.org/idiographic/>, but they are no longer rebuilt during R CMD check. Rebuilding all ten vignettes took 574 seconds on Windows R-devel (68% of the total check time, against CRAN's 10-minute guideline); these three accounted for the large majority of it. The remaining seven vignettes are renumbered 1-7.