fix: suppress SMOTE for multiclass targets (issue #36)

[kimusaku]

Summary

Fixes #36 — SMOTE was being recommended for multiclass classification targets.

Root Cause

_get_target_imbalance_score() in meta_features.py correctly returns 0 when a target has more than 2 classes, but the imbalance score is computed only on the training split. When a rare third class is absent from that split, the method sees only 2 classes and returns a high score, causing SMOTE to be emitted even though the overall task is multiclass (task.is_multiclass = True).

The guard in template_based_adaptation.py (line 307) only blocked multi-target columns (len(task.target_columns) > 1) and did not check task.is_multiclass.

Fix

sapientml_core/adaptation/generation/template_based_adaptation.py

Extend the SMOTE guard to also skip SMOTE when task.is_multiclass is True:

# Before
if "PREPROCESS:Balancing:SMOTE:imblearn" == component.label_name and len(self.task.target_columns) > 1:
    continue

# After
if "PREPROCESS:Balancing:SMOTE:imblearn" == component.label_name and (
    len(self.task.target_columns) > 1 or self.task.is_multiclass
):
    continue

Tests

Added test_smote_not_recommended_for_multiclass to tests/sapientml/test_generatedcode_additional_patterns.py.

The test uses the target_category_binary_imbalance column (imbalance score ≈ 0.913, which would normally trigger SMOTE) but forces task.is_multiclass = True to simulate the edge case where the full dataset has a 3rd rare class absent from the training split. It asserts that "SMOTE" does not appear in the generated code.

Existing binary SMOTE test (test_additional_classifier_works_with_preprocess) continues to pass, confirming no regression for the binary case.