residuals.rglmb() / residuals.rlmb() / residuals.summary.rglmb():
Fixed the ysim argument to substitute for the observed response y
(fitted value held fixed at each draw), not the reverse. The previous
behavior substituted ysim for the fitted value while holding y fixed,
which does not correspond to a standard posterior-predictive residual check
and contradicted this function's own ysim ("simulated responses")
documentation. The corrected convention matches glmbayes's
residuals.glmb() (simulate new response data, recompute residuals against
the same fitted values, and compare to the actual residuals for outlier
diagnostics) — the two packages' ysim semantics were inconsistent since
this function was added post-split; they are now aligned. Behavior for the
default case (ysim = NULL) is unchanged.
Configure (Linux/macOS): -DUSE_OPENCL is set only when a non-PoCL
OpenCL platform exposes at least one GPU device (same policy as
glmbayes 0.9.75). PoCL-only ICD stacks (typical on CRAN debian-gcc
incoming) no longer enable OpenCL at compile time, so kernel builds do not
touch ~/.cache/pocl (avoids incoming NOTE on tempfile_* there). Real GPU
installs (NVIDIA/AMD/Intel) are unchanged. configure.win is unaffected
(Windows never ran a runtime OpenCL probe).
Configure policy: Removed tools/rcpp_include.R / tools/patch_rcpp_function_h.R
and the glmbayes_getRegisteredNamespace compatibility shim (src/glmbayes_getRegisteredNamespace.{cpp,h})
from configure and configure.win. Builds now rely on standard
LinkingTo: Rcpp and the CRAN Rcpp (>= 1.1.1) requirement instead of
probing/patching Rcpp/Function.h or recommending a GitHub install of Rcpp
(same policy fix as glmbayes 0.9.73/0.9.75).
CI: Replaced .github/workflows/rhub.yaml with the current
glmbayes version, removing dead matrix/report logic that referenced the
now-removed tools/patch_rcpp_function_h.R script.
Windows builds now link against the TBB libraries provided by
RcppParallel (RcppParallel::RcppParallelLibs() and
-DRCPP_PARALLEL_USE_TBB=1 in configure.win). This fixes undefined
reference errors to tbb::detail::r1::wait_on_address /
notify_by_address_one seen on current Windows toolchains
(win-builder / R-universe).
Cap RcppParallel to 2 threads in the pfamily, Prior_Setup, and
simfuncs examples (\dontshow{setThreadOptions(numThreads = 2)})
so CRAN checks do not report “CPU time > 2.5 times elapsed time”.
Default parallel sampling for users is unchanged.
First CRAN release of glmbayesCore. This package provides the compiled
sampling engine and related “core” R interfaces (prior families, matrix-level
samplers such as rglmb / rlmb, envelope construction, and optional OpenCL
acceleration) that previously lived inside glmbayes and are now packaged
separately for reuse.
End users of Bayesian GLM/LM modelling should continue to use glmbayes for the formula interface and S3 methods. glmbayesCore is the developer / backend layer those packages build on.
Future work includes pointing glmbayes (and other downstream packages) at this Core package as their shared sampling engine, so that the iid envelope stack is maintained in one place.
Version 0.5.2 (rather than 0.1.0) reflects that the engine API is already substantially complete from its prior life inside glmbayes, while leaving room for integration-driven releases before a 1.0.0 freeze.