[Hanabi9248]

Feature.evaluate accepts numeric assignment operands, including fractions.Fraction, but serializes each binary operation without grouping the operands. For Feature('x/y') with x=Fraction(1, 2) and y=Fraction(1, 3), this builds 1/2/1/3 and returns 1/6 instead of 1.5. The wrong value also changes real Predicate decisions.

Parenthesize each serialized operand before the existing evaluation. This preserves operand order and the existing validation, categorical features, Decimal/constant compatibility and final float conversion. It does not change the unary grammar or claim exact rational arithmetic or complete arbitrary Number support.

Validation against master 0ce916f:

• Python 3.11.16: nine new real Feature/Predicate regression methods produce four assertion failures and five passing controls on the baseline; all nine pass with this one-line change.
• The separate, unchanged upstream tests/rule_induction/trxf/core/test_feature.py suite passes all eight methods on Python 3.11.16.
• Python 3.14.1: the new nine-method regression has the same baseline/fixed result. The separate existing eight-method Feature suite has seven passes and one error: its malformed-parenthesis case expects tokenize.TokenError but reaches IndexError in the unchanged parser. That failure is retained and is outside this arithmetic fix.
• The tests execute complete real Feature, Predicate and BooleanEvaluator modules. Namespace-only package containers avoid unrelated optional top-level imports; no numeric operations, tokenizer, eval or predicate results are mocked. Full installed-package imports and the full repository suite were not run.
• Native Windows execution used one CPU, 768 MiB aggregate memory, 16-process limit, private temporary files and single-threaded BLAS settings. No dependencies were installed. Source and added-test whitespace checks passed.

Run the added regression with: python tests/rule_induction/trxf/core/test_feature_numeric_operands.py -v

This is separate from #155's operand-pop order correction and #211's metrics changes. AI assistance was used to prepare and review the patch; the reported test results come from actual local runs.