NEWS

oda 0.1.6

Directional UniODA LOO for binary attributes (direction = "greater" / "less") now refits each fold with the declared direction. The binary 2x2-table shortcut re-selected the better direction per fold, so when the data ran against the declared direction LOO scored the reversed rule (ess_loo = -ess, a spurious loo_not_stable, and a LOO p for the opposite rule). 0.1.4 fixed the same issue for ordered attributes at the call site; both count-table shortcuts now refuse directional rules themselves, including the early eval_order = "loo_then_mc" stability gate.
oda_loo_for_rule() now stops if the rule's direction contradicts the declared direction, and stops if directional binary LOO predictions differ from the declared-direction rule (an internal invariant).
This also applied to weighted fits with priors_on = FALSE, which used the same 2x2 shortcut. Non-directional fits, weighted fits with priors on (which already refit every fold), and CTA are unchanged.

Directional UniODA Monte Carlo p-values (direction = "greater" / "less") now score each permutation exactly as a full refit does. The fast scorer received the public direction label, so "greater" permutations were scored with the best "1->0" cut; it also scored inadmissible cuts, which changed p when the observed ESS was <= 0. Non-directional MC p-values are unchanged (for "off" the admissible maximum equals max |ESS| over all cuts).
Weighted UniODA (non-uniform case weights, priors on) now matches MegaODA.exe: the rule maximises the count-priors objective sum(w x correct) / n_class over raw class counts, as CTA already does, instead of WESS; the MC p permutes (class, weight) pairs jointly with that objective as the statistic; and LOO refits every fold (the count-table shortcuts re-select folds by ESS). Reported ESS is still WESS. Unweighted and priors-off fits are unchanged. Myeloma V4 now gives MegaODA's cut 890.25 (was 64.0), WESS 6.93, LOO stable.

Weighted CTA Monte Carlo p-values now match CTA.exe: each permutation shuffles the attribute relative to the observations (class and case weight stay together) and the test statistic is the count-priors objective that selects the cut, not WESS. Reproduces CTA.exe significance for all 10 weighted myeloma root attributes. Unweighted CTA is unchanged.
LOO refits each fold with the same optimizer that chose the training rule. UniODA: tied folds are refit (MAXSENS / SAMPLEREP) instead of taking the first-identified cut, and directional rules always refit (the shortcut ignored the direction); matches MegaODA LOO tables. CTA: node LOO refits each fold with the CTA scan (.cta_loo_scan()) instead of generic ODA; matches CTA.exe (cta_demo MODEL8 node 6 STABLE; myeloma node 2 V4 UNSTABLE).
CTA fits every numeric attribute with the CTA scan, with or without case weights (uniform weights reduce its objective to ESS), so each node uses one optimizer for the cut, the MC p and LOO. Categorical attributes still use generic ODA.

Weighted ordered CTA splits now match CTA.exe. .cta_ordered_scan() and .cta_fast_scan_perm() choose the cut that maximises count-priors weighted accuracy (sum of correct case weights divided by the raw class count) among admissible cuts, instead of the outermost cut with sensitivity > 0.5. This reproduces all 118 attribute WESS reports in the CTA.exe myeloma MINDENOM=1 output (#19).
mindenom is now a gate on the selected cut, as in CTA.exe: a split whose smaller side has fewer than mindenom observations is rejected rather than replaced by a different cut. This stops ESS rising as mindenom increases.
Reported ESS is always the weighted ESS (class weight totals rescaled to 1), including with priors_on = FALSE, so it stays within [-100, 100] and does not depend on the scale of the weights (#20).
The Monte Carlo permutation scan shares one implementation with the observed scan; the mindenom gate is not applied to permutations.
Weighted numeric binary (0/1) attributes now use the same CTA scan, so their direction (and WESS, which can be negative) matches CTA.exe.
The weighted CTA path now drops observations missing on the attribute before the scan, Monte Carlo permutations and LOO. Previously LOO scored missing rows through the rule, which understated LOO WESS for attributes with missing values.

Added missing return-value documentation for SDA print and summary methods, as requested by CRAN.
Replaced inappropriate \dontrun{} examples with \donttest{} or runnable examples, as requested by CRAN.
Audited exported functions and registered S3 methods for missing return-value documentation.

Fixed oda_fit(mcarlo = N) so numeric mcarlo values enable Monte Carlo and set the iteration cap.
Preserved rule, confusion, training metrics, and LOO details when a fitted model does not pass a LOO gate.
Improved print() and summary() for ODA fits so rule-bearing models show the learned rule, training evidence, and LOO evidence.
Enabled unweighted multiclass / multicategorical LOO reporting when loo = "on" is requested.
Clarified that multicategorical LOO reports LOO confusion and ESS but does not report a LOO p-value because there is no canon-aligned C x C Fisher p-value.
Clarified p-value directionality: MC permutation p-values are two-tailed; binary LOO Fisher p-values are one-tailed (MPE hold-out canon). Output labels now reflect the distinction.
Fixed nominal multiclass rule display.
Fixed multiclass summary output so class-level LOO PAC is visible.
Fixed PAC percentage formatting to avoid double-scaling percent values.
Added myeloma and cta_demo as package datasets for public examples.
Re-enabled top-level vignette builds.
Updated protein and migraine vignettes to reflect corrected LOO, rule, and p-value behavior.
Kept oda_sample_size() runnable example meaningful while reducing CRAN example runtime.

Added fixture tests for directional categorical LOO: binary fixed direction_map (LOO ESS, Fisher p one-tailed, p < 0.001) and multiclass direction = "ascending" (LOO ESS, confusion, no LOO Fisher p, print states "not reported").
Corrected directional-oda article: multiclass categorical LOO is supported with loo = "on"; clarified that MC p and LOO p are separate calculations.
Fixed binary and categorical rule display - was showing <categorical/binary rule> placeholder. Now shows actual level-to-class mappings, e.g. {low} --> 0 | {high} --> 1.
Added cta_confusion_matrix(tree) convenience wrapper that returns the 2x2 integer training confusion matrix directly from a cta_tree (previously required two-step cta_confusion_table() + as_confusion_matrix() call).
Added attr_names length guard in oda_cta_fit() - error when supplied names do not match number of attributes.
Improved as_cta_candidates() error message when X argument is omitted.
Added regression tests for multiclass / multicategorical LOO reporting.
Added tests for summary/print LOO evidence.
Added scope guardrail tests for LORT lean-fit invariants and SORT/GORT export absence.
Added CONTRIBUTING.md checklist for recursive CTA scope and release checks.