[topepo]

This PR increases the speed of extracting rules. The main cause was that extract_rules.party() looked up each leaf's parent by looping over every node in the tree, which is quadratic or worse per tree. Several other methods had the same per-leaf *_find_parent pattern. Most users extract rules from every tree in an ensemble, so the work was benchmarked on that case.

Changes

Speedups

• party, cforest, ranger, randomForest, grf: extract_rules.party() now does one top-down walk per tree that carries the split expressions down and emits a rule at each leaf. ranger, randomForest and grf also skipped a row-by-row computation of fitted node ids (nrow × depth per tree) by passing a zero-row copy of the data.
• rpart: one top-down walk with a precomputed split-row lookup.
• xgboost, lightgbm, bart, catboost: one model dump, getTrees() call or JSON export now covers all requested trees. Previously it was repeated per tree, which made the total quadratic. Each tree is then converted with a top-down walk.
• aorsf (ObliqueForest): the slow R6 accessors (get_means(), get_stdev(), get_fctr_info(), get_names_x()) are cached once, and each tree is a top-down walk.
• C5.0 and Cubist: already fast, unchanged apart from the shared checks below.

Output is identical() to the old code wherever the old code could produce output.

Timings

Method	Before	After	Case
ranger	115 s	1.7 s per 20 trees	2 small trees before, 20 trees on 2,000 rows after
randomForest	422 s	1.7 s per 20 trees	2 small trees before, 20 trees on 2,000 rows after
grf	661 s	0.25 s	10 trees, 1,000 rows
rpart	4.1 s	0.12 s
bart	7.6 s	0.19 s	100 trees × 2 chains
lightgbm	6.8 s	0.25 s	50 trees
xgboost	1.2 s	0.19 s	40 trees
catboost	0.6–1.1 s	0.17 s	100 trees
aorsf	1.5 s	0.39 s	20 trees

The ranger and randomForest "before" and "after" rows are not the same workload, so they understate the improvement.

Test suite timings

File	Before	After
test-cforest.R	skipped (estimated 1–2 hours if enabled)	about 2 s
test-party.R	skipped (7:43 when enabled)	about 1.7 s

Both files were slow because of model fitting, not rule extraction. partykit::cforest() and ctree() were fit on wa_trees, where county has 30 levels (one 3-tree cforest fit took about 170 s, while extraction took 0.01–0.03 s). A new test helper, get_wa_trees_small_data(), keeps 200 rows and the 3 most common counties and labels the rest other. A 3-tree fit now takes 0.02 s. The skip() and the random skip_if(runif(1) <= 0.1) are removed, and tests/testthat/_snaps/cforest.md is new because those snapshots were never recorded.

API and behavior changes

• tree accepts a vector for ObliqueForest, bart, catboost, lightgbm and xgboost, matching the other ensemble methods.
• Shared validation: new internal check_tree_type() and check_tree_range() are used by every method with a tree argument (including C5.0 trials, Cubist committees, cforest and active_predictors.cforest()). The range error now states the upper limit, for example "tree must be between 1 and 3, the number of trees in the model."
• Tree counts: num_ensemble_members() is the single source for the number of trees. Internal lgb_num_trees(), xgb_num_trees() and bart_num_trees() take an already-read dump so the model is not dumped a second time.
• rlang::check_dots_empty() is now called first in every extract_rules() method, so typos such as trees = 2 error instead of being silently ignored. As a result data is no longer accepted by extract_rules.C5.0().
• grf: the fallback of 1000 trees is removed, num_ensemble_members.grf() aborts clearly when the number of trees is missing, and matrices without column names no longer error.

Tests

• New test that every training row matches exactly one party rule and that the rule id equals predict(type = "node").
• New "handles several trees in one call" tests for lightgbm, xgboost, catboost, ObliqueForest and bart.
• New tests with snapshots for check_tree_type() and check_tree_range().
• Snapshots updated for the new error messages. The tree = c(1, 2) error snapshots were removed because that call is now valid. data = was removed from the C5.0 tests.
• Full suite: 2,545 passing, 0 failing, 1 skipped (aorsf "On Mac").