e2tree

• **CatBoost loss-function parsing**: in some catboost releases the loss_function returned by catboost.get_model_params() is a list (e.g. list(type = "MultiClass", border = 0.5)) instead of a bare character scalar. The previous detection code piped this list directly into startsWith() and failed with Error: non-character object(s). The new .catboost_extract_loss() canonicalises any of the documented shapes (plain character, list-with-type, list-with-name, list whose first element holds the name) to a length-one character string before classification/regression dispatch and falls back to attr(ensemble, "params") when catboost.get_model_params() does not expose the loss.
• **CatBoost 1.2.x compatibility**: starting from catboost 1.2 the R objects returned by catboost.train() carry the class catboost.Model instead of the previous catboost.CatBoost. All three S3 generics in the adapter layer (get_ensemble_type, extract_terminal_nodes, get_ensemble_predictions) now register methods for both class names, sharing a single internal worker, so models trained with either version are accepted out of the box. ensemble_backend() and .SUPPORTED_BACKENDS recognise both class names. createDisMatrix() now sets the e2tree_label attribute on CatBoost ensembles automatically when the user passes label, eliminating the previously required manual attr(ensemble, "e2tree_label") <- ... step.
• **moda() accepts non-factor responses**: classification ensembles trained on integer 0/1 responses (typical of gbm with the bernoulli distribution) no longer trigger Error in res[1] <- as.character(levels(x)[res[1]]) : replacement has length zero. The internal moda() helper now coerces the modal value to character independently of whether the response is a factor, an integer, a character or a numeric vector. Affected call sites: eStoppingRules(), e2tree(), vimp().
• **Vignette models.Rmd corrected**: the regression examples now pass the response column through data together with label = "<response_name>", as required by createDisMatrix() for regression backends; the gbm regression example sets n.minobsinnode = 2 and bag.fraction = 0.8 so it does not fail the gbm internal sample-size check on the small mtcars training set; the lightgbm regression example uses num_leaves = 8 and min_data_in_leaf = 2, which prevents the small-sample degenerate case caught by validate_terminal_nodes(). The classification example based on the binary is_setosa indicator now passes a factor copy of the response to e2tree() while leaving the numeric copy to gbm.

Multi-backend hardening (2026-04)

• **Adapter contract validation**: createDisMatrix() now calls validate_terminal_nodes() immediately after extract_terminal_nodes(). Malformed adapter output (wrong row count, non-numeric leaf columns, or the all-same-leaf degenerate case that previously made GBM silently produce an all-zero D) is now rejected with an informative error.
• **Backend tagging**: dissimilarity matrices returned by createDisMatrix() carry a new ensemble_backend attribute identifying the backend used. e2tree() and as.rpart() warn when the supplied ensemble does not match this tag, catching the common "stale D reused with a different model" mistake.
• **xgboost 3.x compatibility**: models built via the high-level xgboost::xgboost() wrapper (which adds an xgboost class on top of xgb.Booster) are now handled correctly; previously the wrapper's predict.xgboost method rejected xgb.DMatrix input and broke leaf extraction.
• **Chunked dissimilarity path wired in**: when chunk_size < n_obs the function now actually uses compute_dissimilarity_chunked() (previously the parameter was accepted but ignored). Optional use_disk = TRUE persists the result via saveRDS().
• **C++ type checks**: compute_cooccurrences_cpp(), compute_cooccurrences_sparse_cpp() and compute_all_cooccurrences_cpp() now validate that Tree* and resp columns are integer/double, raising a clear error rather than silently coercing factors/characters to zeros.
• **Adapter de-duplication**: the three S3 generics share a single .unsupported_backend() helper and a single supported-classes list; the xgboost adapters share a .xgb_dmatrix() builder.
• **Documentation**: ?createDisMatrix now contains an "Interpretation note (RF vs boosting)" section explaining why the dissimilarity scale differs systematically between bagging and boosting backends.
• **Tests**: new tests/testthat/test-multi-backend.R exercises the D invariants (square, zero diagonal, symmetric, non-degenerate, correctly tagged) for each available backend, and covers the validator and the default-method failure paths.

S3 class system overhaul

• e2tree class now properly listed as first in the class vector (c("e2tree", "list")).
• New S3 methods for e2tree: predict(), fitted(), residuals().
• predict.e2tree() replaces ePredTree() as the standard prediction interface. For regression, returns a data frame with fit and sd (node-level standard deviation). For classification, returns a data frame with fit, accuracy, and score.
• fitted.e2tree() returns fitted values for training data.
• residuals.e2tree() returns residuals for regression E2Trees.
• methods(class = "e2tree") now shows: as.rpart, e2splits, fitted, nodes, plot, predict, print, residuals, summary.
• methods(class = "eValidation") now shows: measures, plot, print, proximity, summary.

Accessor functions (new)

• nodes(): Extract tree node data frame from an e2tree object, with optional terminal filter.
• e2splits(): Extract split and categorical split information.
• measures(): Extract validation measures from an eValidation object.
• proximity(): Extract proximity matrices (ensemble, e2tree, or both) from an eValidation object.

Coercion methods (new)

• as.rpart(): Generic and method for converting e2tree to rpart format.
• as.party(): Method for converting e2tree to partykit's constparty format (registered conditionally when partykit is installed). Produces proper bar plots in terminal nodes for classification trees.
• rpart2Tree() retained for backward compatibility with a deprecation note.

Validation framework improvements

• eValidation() gains a test argument: "mantel" (Mantel test only), "measures" (divergence/similarity measures only), or "both" (default). This allows choosing between association testing and agreement testing.
• print.eValidation(), summary.eValidation(), and plot.eValidation() updated to handle all three test modes gracefully.

Variable importance improvements

• vimp(): Auto-detects classification/regression from the e2tree object; type argument now optional.
• Fixed incorrect y-axis label ("Variance" instead of "Variable") in variable importance plots.
• Regression variable importance bars now sorted by importance (previously unsorted).
• Consistent column naming (Variable, MeanImpurityDecrease) across classification and regression.
• Internal logic refactored into .vimp_classification() and .vimp_regression().

Documentation

• All man page titles standardized to Title Case.
• \dontrun{} replaced with \donttest{} in all examples; interactive-only examples wrapped in if (interactive()).
• Examples updated to use accessor functions and predict() instead of direct $ access.
• New vignette e2tree-introduction covering classification, regression, validation (Mantel test, divergence measures, LoI decomposition), and comparison with partykit/stablelearner.
• ePredTree() documentation updated with deprecation note pointing to predict.e2tree().

Package infrastructure

• Added partykit, knitr, rmarkdown to Suggests.
• Added VignetteBuilder: knitr to DESCRIPTION.
• e2tree object now stores fitted.values, y, and data for S3 method support.
• Conditional .onLoad hook for registering as.party.e2tree when partykit is loaded.