C++ revamp + add GGM

.editorconfig

@@ -0,0 +1,23 @@
+root = true
+
+[*]
+end_of_line = lf
+insert_final_newline = true
+trim_trailing_whitespace = true
+charset = utf-8
+
+[*.{R,r}]
+indent_style = space
+indent_size = 2
+
+[*.{h,cpp}]
+indent_style = space
+indent_size = 4
+
+[*.{Rmd,md}]
+indent_style = space
+indent_size = 2
+trim_trailing_whitespace = false
+
+[Makevars*]
+indent_style = tab

.github/copilot-instructions.md

@@ -0,0 +1,94 @@
+# bgms — AI Agent Instructions
+
+Rules for AI agents (Copilot, Claude, etc.) working on this codebase.
+
+## R code style
+
+- Use `=` for assignment, not `<-`.
+- **Exception**: inside function-call arguments where the result must
+  be captured, use `<-`. With `=` R treats it as a named argument.
+  Example: `expect_message(result <- foo(), "pattern")`.
+- No space between `if`/`for`/`while` and `(`: write `if(`, not `if (`.
+- Enforced by `inst/styler/bgms_style.R`. Run before committing:
+  ```r
+  source("inst/styler/bgms_style.R")
+  styler::style_pkg(style = bgms_style)
+  ```
+  After running, check test files for `expect_*(result = ...)` and
+  revert those to `result <- ...`.
+
+## Exported R functions (Tier 1)
+
+- Full roxygen2 block: `@title`, `@description`, `@param` (all params),
+  `@return`, `@examples`, `@seealso`, `@export`.
+- `@param` for every parameter, no exceptions. Use `@inheritParams bgm`
+  for shared parameters.
+- `@return` must describe the structure, not just the type.
+- `@examples` must be runnable by `R CMD check`. Use `\donttest{}` for
+  slow examples. Use `\dontrun{}` only when the example truly cannot
+  run in a clean R session.
+- Add `@family` tags: `model-fitting`, `posterior-methods`, `prediction`,
+  `extractors`, `diagnostics`.
+- Never use `@keywords internal` on exported functions.
+
+## Internal R functions (Tier 2)
+
+- Use plain `#` comments, not `#'` roxygen. No `@noRd` unless the
+  function genuinely benefits from `@inheritParams`.
+- Use `# ====` banners for file sections, `# ----` banners per function.
+- Format:
+  ```r
+  # ------------------------------------------------------------------
+  # function_name
+  # ------------------------------------------------------------------
+  # One-sentence purpose.
+  #
+  # @param arg1  What it is.
+  # @param arg2  What it is.
+  #
+  # Returns: What comes back.
+  # ------------------------------------------------------------------
+  ```
+
+## C++ headers (Tier 3)
+
+- Doxygen `/** */` blocks on all classes, public methods, and free
+  functions. Use `///` for struct fields and enum values.
+- Class-level doc is mandatory: explain purpose and place in the
+  `BaseModel → GGMModel/OMRFModel` hierarchy.
+- Private methods get `/** */` blocks. Prefer over-documenting.
+- For ported algorithms, add an origin note (source, algorithm name).
+- No Doxygen in `.cpp` files — use inline `//` comments only.
+
+## C++ implementations (Tier 4)
+
+- Inline `//` comments for non-obvious steps only.
+- Use `// --- Phase N: ... ---` section comments in long functions.
+- Reference formulas by name or equation number.
+
+## Error messages
+
+- State what went wrong, why, and what the user can do about it.
+- Use complete sentences. Do not start with "Error:" — R adds that.
+- See `R/validate_data.R` as the exemplar.
+
+## Commit conventions
+
+- One commit per audit task. Message format:
+  `docs: short description (audit #N)`.
+- When modifying a function signature, update its documentation in
+  the same commit.
+- When adding a new exported function, add it to `_pkgdown.yml`.
+
+## Do not
+
+- Add `@keywords internal` to exported functions.
+- Use `\dontrun{}` when `\donttest{}` suffices.
+- Add Doxygen blocks in `.cpp` implementation files.
+- Use `<-` for assignment in R code.
+- Use AI-style prose: no superlatives, no hedging qualifiers, no
+  transition phrases ("Furthermore", "Moreover", "Additionally").
+- Write session-oriented documentation ("in your R session",
+  "you should see", "run this in your console"). Documentation
+  describes what a function does and returns, not what the reader
+  should do interactively.

.github/workflows/lint.yaml

@@ -0,0 +1,34 @@
+# Workflow derived from https://github.com/r-lib/actions/tree/v2/examples
+on:
+  push:
+    branches: [main]
+  pull_request:
+    branches: [main]
+
+name: lint
+
+permissions: read-all
+
+jobs:
+  lint:
+    runs-on: ubuntu-latest
+    name: lint
+    env:
+      GITHUB_PAT: ${{ secrets.GITHUB_TOKEN }}
+    steps:
+      - uses: actions/checkout@v5
+
+      - uses: r-lib/actions/setup-r@v2
+        with:
+          use-public-rspm: true
+
+      - uses: r-lib/actions/setup-r-dependencies@v2
+        with:
+          extra-packages: any::lintr, local::.
+          needs: lint
+
+      - name: Lint
+        run: lintr::lint_package()
+        shell: Rscript {0}
+        env:
+          LINTR_ERROR_ON_LINT: true

.gitignore

@@ -18,3 +18,4 @@ src/sources.mk
 docs/*
 /doc/
 /inst/doc/
+dev/plans/

.lintr

@@ -0,0 +1,20 @@
+linters: linters_with_defaults(
+    assignment_linter = NULL,
+    indentation_linter = NULL,
+    brace_linter = NULL,
+    spaces_inside_linter = NULL,
+    spaces_left_parentheses_linter = NULL,
+    commented_code_linter = NULL,
+    trailing_blank_lines_linter = NULL,
+    object_name_linter = NULL,
+    object_length_linter = NULL,
+    object_usage_linter = NULL,
+    return_linter = NULL,
+    line_length_linter = NULL,
+    T_and_F_symbol_linter(),
+    seq_linter()
+  )
+exclusions: list(
+    "R/RcppExports.R",
+    "inst/"
+  )

CONTRIBUTING.md

@@ -0,0 +1,72 @@
+# Contributing to bgms
+
+Thank you for your interest in contributing to bgms. This document
+covers the practical steps for building, testing, and submitting
+changes.
+
+## Getting started
+
+```bash
+git clone https://github.com/Bayesian-Graphical-Modelling-Lab/bgms.git
+cd bgms
+```
+
+Build and check the package:
+
+```r
+devtools::document()
+devtools::check()
+```
+
+## Code style
+
+bgms uses `=` for assignment (not `<-`) and omits the space between
+`if`/`for`/`while` and `(`. These rules are enforced by the custom
+styler in `inst/styler/bgms_style.R`. Run before committing:
+
+```r
+source("inst/styler/bgms_style.R")
+styler::style_pkg(style = bgms_style)
+```
+
+## Documentation
+
+- **Exported R functions** get full roxygen2 blocks (`@title`,
+  `@description`, `@param`, `@return`, `@examples`, `@seealso`,
+  `@export`).
+- **Internal R functions** use plain `#` comment blocks, not `#'`
+  roxygen.
+- **C++ headers** use Doxygen `/** */` blocks on classes and methods.
+- **C++ implementations** use inline `//` comments only.
+
+When modifying a function signature, update its documentation in the
+same commit. When adding a new exported function, add it to
+`_pkgdown.yml`.
+
+## Tests
+
+Run the test suite with:
+
+```r
+devtools::test()
+```
+
+Test files live in `tests/testthat/`. Each `test_that()` description
+should be an imperative sentence.
+
+## CI
+
+Pull requests are checked automatically by GitHub Actions:
+
+- `R-CMD-check.yaml` — builds and checks on multiple platforms
+- `test-coverage.yaml` — reports test coverage
+- `pkgdown.yaml` — builds the documentation site
+
+## Commit messages
+
+Use the format: `type: short description`. Common types: `feat`,
+`fix`, `docs`, `refactor`, `test`.
+
+## Contributors
+
+All contributors are acknowledged in `inst/CONTRIBUTORS.md`.

DESCRIPTION

@@ -1,6 +1,6 @@
 Package: bgms
 Type: Package
-Title: Bayesian Analysis of Networks of Binary and/or Ordinal Variables
+Title: Bayesian Analysis of Graphical Models
 Version: 0.1.6.4
 Date: 2026-02-16
 Authors@R: c(
@@ -16,7 +16,10 @@ Authors@R: c(
     person("Gali", "Geller", role = "ctb")
   )
 Maintainer: Maarten Marsman <[email protected]>
-Description: Bayesian variable selection methods for analyzing the structure of a Markov random field model for a network of binary and/or ordinal variables. 
+Description: Bayesian estimation and variable selection for Markov random field
+    models of networks of binary, ordinal, and continuous variables. Supports
+    Gaussian graphical models, multi-group comparison via 'bgmCompare', and
+    provides simulation, prediction, and missing data imputation.
 Copyright: Includes datasets 'ADHD' and 'Boredom', which are licensed under CC-BY 4. See individual data documentation for license and citation.
 License: GPL (>= 2)
 URL: https://Bayesian-Graphical-Modelling-Lab.github.io/bgms/, https://github.com/Bayesian-Graphical-Modelling-Lab/bgms
@@ -45,6 +48,7 @@ Suggests:
     covr,
     ggplot2,
     knitr,
+    MASS,
     parallel,
     qgraph,
     rmarkdown,

NAMESPACE

@@ -23,6 +23,7 @@ S3method(extract_sbm,bgms)
 S3method(predict,bgmCompare)
 S3method(predict,bgms)
 S3method(print,bgmCompare)
+S3method(print,bgm_spec)
 S3method(print,bgms)
 S3method(print,summary.bgmCompare)
 S3method(print,summary.bgms)

NEWS.md

@@ -1,3 +1,15 @@
+# bgms 0.1.6.4
+
+## New features
+
+* Gaussian graphical models (GGM): `bgm(x, variable_type = "continuous")` fits a GGM with Bayesian edge selection. Pairwise effects are partial correlations from the precision matrix.
+* Missing data imputation: `na_action = "impute"` integrates over missing values during MCMC sampling for both ordinal and continuous models.
+
+## Bug fixes
+
+* fixed Blume-Capel centering: `observations_` was stored RAW (0-indexed) while `observations_double_` was CENTERED, causing six downstream sites to use wrong values when `baseline_category != 0`. Fixed by centering `observations_` in the constructor for Blume-Capel variables so both representations are in the same coordinate system.
+* fixed Blume-Capel imputation: zero-category probability had wrong sign and was double-counted. Replaced with unified loop over all categories, matching `simulate_mrf()`.
+
 # bgms 0.1.6.3
 
 ## New features

R/RcppExports.R

@@ -5,8 +5,20 @@ run_bgmCompare_parallel <- function(observations, num_groups, counts_per_categor
     .Call(`_bgms_run_bgmCompare_parallel`, observations, num_groups, counts_per_category, blume_capel_stats, pairwise_stats, num_categories, main_alpha, main_beta, pairwise_scale, pairwise_scaling_factors, difference_scale, difference_selection_alpha, difference_selection_beta, difference_prior, iter, warmup, na_impute, missing_data_indices, is_ordinal_variable, baseline_category, difference_selection, main_difference_selection, main_effect_indices, pairwise_effect_indices, target_accept, nuts_max_depth, learn_mass_matrix, projection, group_membership, group_indices, interaction_index_matrix, inclusion_probability, num_chains, nThreads, seed, update_method, hmc_num_leapfrogs, progress_type)
 }
 
-run_bgm_parallel <- function(observations, num_categories, pairwise_scale, edge_prior, inclusion_probability, beta_bernoulli_alpha, beta_bernoulli_beta, beta_bernoulli_alpha_between, beta_bernoulli_beta_between, dirichlet_alpha, lambda, interaction_index_matrix, iter, warmup, counts_per_category, blume_capel_stats, main_alpha, main_beta, na_impute, missing_index, is_ordinal_variable, baseline_category, edge_selection, update_method, pairwise_effect_indices, target_accept, pairwise_stats, hmc_num_leapfrogs, nuts_max_depth, learn_mass_matrix, num_chains, nThreads, seed, progress_type, pairwise_scaling_factors) {
-    .Call(`_bgms_run_bgm_parallel`, observations, num_categories, pairwise_scale, edge_prior, inclusion_probability, beta_bernoulli_alpha, beta_bernoulli_beta, beta_bernoulli_alpha_between, beta_bernoulli_beta_between, dirichlet_alpha, lambda, interaction_index_matrix, iter, warmup, counts_per_category, blume_capel_stats, main_alpha, main_beta, na_impute, missing_index, is_ordinal_variable, baseline_category, edge_selection, update_method, pairwise_effect_indices, target_accept, pairwise_stats, hmc_num_leapfrogs, nuts_max_depth, learn_mass_matrix, num_chains, nThreads, seed, progress_type, pairwise_scaling_factors)
+get_explog_switch <- function() {
+    .Call(`_bgms_get_explog_switch`)
+}
+
+rcpp_ieee754_exp <- function(x) {
+    .Call(`_bgms_rcpp_ieee754_exp`, x)
+}
+
+rcpp_ieee754_log <- function(x) {
+    .Call(`_bgms_rcpp_ieee754_log`, x)
+}
+
+compute_conditional_ggm <- function(observations, predict_vars, precision) {
+    .Call(`_bgms_compute_conditional_ggm`, observations, predict_vars, precision)
 }
 
 compute_conditional_probs <- function(observations, predict_vars, pairwise, main, num_categories, variable_type, baseline_category) {
@@ -21,10 +33,26 @@ sample_bcomrf_gibbs <- function(num_states, num_variables, num_categories, pairw
     .Call(`_bgms_sample_bcomrf_gibbs`, num_states, num_variables, num_categories, pairwise, main, variable_type_r, baseline_category, iter, seed)
 }
 
+sample_ggm_direct <- function(num_states, precision, means, seed) {
+    .Call(`_bgms_sample_ggm_direct`, num_states, precision, means, seed)
+}
+
 run_simulation_parallel <- function(pairwise_samples, main_samples, draw_indices, num_states, num_variables, num_categories, variable_type_r, baseline_category, iter, nThreads, seed, progress_type) {
     .Call(`_bgms_run_simulation_parallel`, pairwise_samples, main_samples, draw_indices, num_states, num_variables, num_categories, variable_type_r, baseline_category, iter, nThreads, seed, progress_type)
 }
 
+run_ggm_simulation_parallel <- function(pairwise_samples, main_samples, draw_indices, num_states, num_variables, means, nThreads, seed, progress_type) {
+    .Call(`_bgms_run_ggm_simulation_parallel`, pairwise_samples, main_samples, draw_indices, num_states, num_variables, means, nThreads, seed, progress_type)
+}
+
+sample_ggm <- function(inputFromR, prior_inclusion_prob, initial_edge_indicators, no_iter, no_warmup, no_chains, edge_selection, seed, no_threads, progress_type, edge_prior = "Bernoulli", beta_bernoulli_alpha = 1.0, beta_bernoulli_beta = 1.0, beta_bernoulli_alpha_between = 1.0, beta_bernoulli_beta_between = 1.0, dirichlet_alpha = 1.0, lambda = 1.0, na_impute = FALSE, missing_index_nullable = NULL) {
+    .Call(`_bgms_sample_ggm`, inputFromR, prior_inclusion_prob, initial_edge_indicators, no_iter, no_warmup, no_chains, edge_selection, seed, no_threads, progress_type, edge_prior, beta_bernoulli_alpha, beta_bernoulli_beta, beta_bernoulli_alpha_between, beta_bernoulli_beta_between, dirichlet_alpha, lambda, na_impute, missing_index_nullable)
+}
+
+sample_omrf <- function(inputFromR, prior_inclusion_prob, initial_edge_indicators, no_iter, no_warmup, no_chains, edge_selection, sampler_type, seed, no_threads, progress_type, edge_prior = "Bernoulli", na_impute = FALSE, missing_index_nullable = NULL, beta_bernoulli_alpha = 1.0, beta_bernoulli_beta = 1.0, beta_bernoulli_alpha_between = 1.0, beta_bernoulli_beta_between = 1.0, dirichlet_alpha = 1.0, lambda = 1.0, target_acceptance = 0.8, max_tree_depth = 10L, num_leapfrogs = 10L, pairwise_scaling_factors_nullable = NULL) {
+    .Call(`_bgms_sample_omrf`, inputFromR, prior_inclusion_prob, initial_edge_indicators, no_iter, no_warmup, no_chains, edge_selection, sampler_type, seed, no_threads, progress_type, edge_prior, na_impute, missing_index_nullable, beta_bernoulli_alpha, beta_bernoulli_beta, beta_bernoulli_alpha_between, beta_bernoulli_beta_between, dirichlet_alpha, lambda, target_acceptance, max_tree_depth, num_leapfrogs, pairwise_scaling_factors_nullable)
+}
+
 compute_Vn_mfm_sbm <- function(num_variables, dirichlet_alpha, t_max, lambda) {
     .Call(`_bgms_compute_Vn_mfm_sbm`, num_variables, dirichlet_alpha, t_max, lambda)
 }