glmbayes 0.9.7

Highlights

• CRAN dependency on nmathopencl: OpenCL statistical kernels are now
  supplied by the nmathopencl package (hard Imports dependency),
  including Windows binaries on CRAN — unblocking GPU/OpenCL builds on Windows
  without vendored nmath in glmbayes.
• Bug fixes (see below): print.lmb() call display, truncated
  dIndependent_Normal_Gamma dispersion prior simulation for
  simulate_prior(), and related documentation/roxygen cleanup.

insight integration

• Added S3 methods for insight accessors on glmb fits
  (model_info, get_parameters, find_parameters, find_algorithm,
  get_data, get_priors). insight moves from Suggests to
  Imports and key generics are re-exported. get_priors() returns
  pfamily(model) (full prior specification, including complete
  Sigma) rather than a marginal-only table.

bayestestR prior-checking integration

• Added simulate_prior(), check_prior(), and describe_prior()
  methods for glmb fits; bayestestR moves from Suggests to
  Imports. Each pfamily() stores a matching pfun (like
  simfun); simulate_prior.glmb() calls it, including the truncated
  inverse-gamma dispersion prior used by dIndependent_Normal_Gamma()
  fits.

Bug fixes

• print.lmb(): Single-response lmb() fits again show a concise
  Call: line (formula, n, data, and related arguments) instead of
  deparsing the full pfamily object, including embedded simfun
  source. When multi-response lmb() / mlmb support was added in
  0.9.6, .mlmb_lmb_display_call() omitted pfamily from the stored
  call for each block fit only; univariate lmb() still used the internal
  .uni_lmb() matched call via do.call(), which inlined the evaluated
  pfamily. Univariate fits now use the same display-call helper as
  mlmb.

Independent Normal-Gamma simulation

• Updates to independent normal gamma simulation to better handle
  non-isotropic priors with highly differentiated implied pweights across
  dimensions.

CPU nmath phase-out

• Removed vendored CPU R Mathlib sources; CPU routines in C++ now use R's
  libR via <Rmath.h>.

OpenCL kernel loading

• OpenCL nmath comes from nmathopencl; glmbayes likelihood/envelope
  kernels remain under inst/cl/, loaded via opencltools.