Meta: validate the randomForest engine (plot/ROC correctness + regression coverage)

[ehrlinger]

Summary

Tracking issue for validating and hardening the randomForest engine paths in ggRandomForests. Historically the package has been exercised mostly via rfsrc; a validation pass across the 6 randomForest-dispatching extractors (gg_rfsrc, gg_error, gg_vimp, gg_variable, gg_roc, calc_roc) × {classification, regression} surfaced real defects with no regression coverage.

Grouped issues

• gg_variable() returns an empty plot for randomForest objects (classification & regression) #80 — gg_variable() returns an empty plot for randomForest objects (classification & regression). Root cause: plot.gg_variable returns a bare list instead of a single ggplot/patchwork for the randomForest case (same bug class as the v2.7.3 plot.gg_partial*: return a single combined figure instead of a named list #77/fix: plot.gg_partial* return single patchwork figure instead of named list (#77) #78 plot.gg_partial* unification); the gg_variable.randomForest extractor itself is correct.
• gg_roc()/calc_roc() produces a degenerate ROC (~0.5 AUC, 3 points) for randomForest objects #81 — gg_roc()/calc_roc() produce a degenerate ROC (~0.5 AUC, 3 points) for randomForest. Root cause: the curve is built from hard class labels, not OOB vote probabilities; the default which_outcome="all" path warns and silently returns class 1.

Scope (v2.8 cycle, own branch/PR — not the varPro Phase plans)

• Extend tests/testthat/test_plot_layer_data.R with a full randomForest matrix (every supported family × class/regr) using the existing expect_layer_nonempty / expect_layer_has_variation helpers — coverage-first (TDD: gg_variable() returns an empty plot for randomForest objects (classification & regression) #80/gg_roc()/calc_roc() produces a degenerate ROC (~0.5 AUC, 3 points) for randomForest objects #81 fail red first).
• ROC numeric-correctness assertions in test_gg_roc.R (AUC≈1 on separable iris setosa; ROC point count > n_classes; default macro-average curve, no warning).
• Fix gg_variable() returns an empty plot for randomForest objects (classification & regression) #80 (plot.gg_variable → single ggplot/patchwork; do not regress the shared rfsrc path).
• Fix gg_roc()/calc_roc() produces a degenerate ROC (~0.5 AUC, 3 points) for randomForest objects #81 (calc_roc.randomForest uses object$votes OOB probabilities; default = macro-average one-vs-rest). Per-class faceting + ROC CIs remain with Meta: gg_roc enhancements (multi-class + confidence intervals) #72, not here.
• vdiffr snapshots per supported family × type, under test_snapshots.R's existing opt-in guard.
• Minor: plot.gg_rfsrc regression stray labs(colour="Outcome") (Ignoring unknown labels).
• Rd-document gg_partial/gg_survival/gg_brier as intentionally rfsrc-only.

Non-goals

• Per-class / multi-class faceted ROC and ROC confidence intervals — tracked in Meta: gg_roc enhancements (multi-class + confidence intervals) #72.
• Any rfsrc-path behavior change; varPro work.

Closes #80, #81 when complete.

[ehrlinger]

Sibling defect surfaced by the RF coverage-first matrix (validation pass, RF3/RF5): gg_variable(rf, ...) for randomForest classification produces per-variable sub-plots whose geom_smooth cannot compute stat_smooth ('stat_smooth()' requires the following missing aesthetics: x and y when ggplot2::layer_data() is called on the per-variable sub-plot). Reproduction:

set.seed(42)
rf <- randomForest::randomForest(Species ~ ., data = iris)
p  <- plot(gg_variable(rf))
ggplot2::layer_data(p[[1]], 1)   # ERROR

The plot displays via print() (the smooth layer silently drops on multi-class yhat.* columns) but introspection fails. Likely root cause: plot.gg_variable builds per-variable geom_smooth over a reshaped multi-class long-form data frame without ensuring an aes(x = ..., y = ...) is mapped on the smooth layer for each class. Out of scope for the current PR (#80/#81/wart) — logging here as a follow-up to consider in a later randomForest classification gg_variable cleanup pass.