Fixed an issue with standardization. Variables with zero variance get turned into Inf. What we do instead is change the variance to Inf, and those variables become all 0 and are given zero coefficients. This problem can exhibit during CV.
Fixed the dense-matrix sp_tmul so that its serial path honors
n_threads. Its n_threads = 1 branch ran a raw Eigen product that
ignored the requested thread count and fell back to Eigen's global
default (all cores), inflating example CPU time to user/elapsed
~10x with no wall-clock gain on a many-core machine (the source of
CRAN's gcc-16 R-devel "CPU time > 2.5 times elapsed" NOTE). The
inst/adelie submodule points at the one-line fix on a forked branch
(bnaras/adelie@zero-variance). Investigation notes and a Docker
reprex live in debug/ (excluded from the package tarball).
grpnet() and cv.grpnet() triggered by
inputs where the BASIL kkt and screen checks disagreed by a
1-ULP gap due to inconsistent multiplication ordering in the bundled
adelie_core solver. As a temporary workaround, the inst/adelie
submodule points at the fix on a forked branch
(bnaras/adelie@fix-kkt-ulp) until upstream pull request
https://github.com/JamesYang007/adelie/pull/159 is merged. A
reprex and root-cause analysis live in bug_reports/kkt_ulp/
(excluded from the package tarball).IOSNPUnphased and IOSNPPhasedAncestry classes.RStateMultiGlm64.NEWS.md file to track changes to the package.