Implemented multiclass classification support across the C++ core and both Python wrappers, including fitting, prediction, metrics, simplification, and visualization.

[gAldeia]

Multiclass support

• Implemented stable, row-wise softmax probabilities.
• Softmax roots now emit one column per class (currently supporting 2–5 classes due to the current signature limit).
• Implemented multiclass log loss, accuracy, and balanced accuracy.
• Added sklearn label encoding/restoration for BrushClassifier and DeapClassifier.
• Added predict_proba support for multiclass programs, individuals, and engines.
• Added C++ and Python multiclass test coverage.
• Added multiclassification.ipynb with sklearn data, tree output, and Graphviz rendering.

Fitting and optimization

• Binary classification now minimizes weighted log loss using sqrt(weight × log_loss) residuals.

• Class weights are propagated from estimator parameters into fitting while preserving existing unweighted behavior.

• Multiclass fitting now optimizes float/logit parameters using weighted multinomial log loss.

• Split thresholds are greedily fitted and refreshed whenever fit() runs:

  • SplitOn fits its decision subtree first.
  • SplitBest refreshes its selected feature and threshold.
  • Boolean branches remain unweighted.

• Split nodes are excluded from Ceres parameter vectors so their thresholds do not interfere with subsequent float parameters.

• Corrected regression split gain to use child-size-weighted variance rather than favoring small partitions.

Simplification

Fixed an issue where multiclass constant simplification treated every subtree as matrix-valued. Class-logit branches are vector-valued, so dispatching Sum as matrix-valued could produce Sum(MatrixF) and incorrectly collapse valid branches.

• Simplifier now evaluates multiclass nodes according to their actual return type.
• Softmax DOT edges are labeled class 0, class 1, etc.
• Added regression coverage for multiclass fitting, simplification, and probability prediction.

Validation

• Python syntax checks passed.
• Notebook JSON/content validation passed.
• C++ extension builds successfully using the generated Makefile (CMake is not available on the current shell's PATH).
• Focused evaluation tests pass.
• All 16 existing regression optimizer tests pass.
• Added coverage for weighted binary fitting, multiclass weighted fitting, split-threshold fitting with float-leaf fitting, and multiclass probability prediction.
• Fixed both previously reported failures:
  • Multiclass fitting no longer attempts to use the scalar-only weight optimizer with an invalid ArrayFJet Softmax callable.
  • predict_proba is now correctly bound for multiclass programs, individuals, and engines.

[gAldeia]

Looks good. approved