Output cleaning coerces boolean metric fields to integer 0/1

Summary

The scaffold's _clean() helper checks integer types before preserving booleans. In Python, bool is a subclass of int, so ordinary boolean metric fields are converted to 1 and 0 when dataframe records are prepared for JSON output.

This affects intervention per-instance fields such as eligible and eligible_wald_sel, which are boolean conditions in the metric API but can be serialized as numeric values in summary.json.

The numeric representation is truthy-equivalent, but it changes the output schema and makes audit/consumer code less reliable because a logical eligibility flag is indistinguishable by type from an integer metric or count.

Expected behavior

Boolean and NumPy boolean values should remain JSON booleans (true/false). Integer and NumPy integer values should continue to serialize as ordinary Python integers.

Suggested fix

Handle (bool, np.bool_) before the integer conversion in _clean() and add focused regression coverage for scalar cleaning and dataframe record conversion.

Start at the scaffold's _clean() helper and trace the dataframe record conversion path. Add focused regression coverage for scalar cleaning and dataframe records, confirming Python and NumPy booleans remain JSON booleans while integer values remain ordinary Python integers.