Preserve bool dtype in SingleAmountTaxScale.calc()

[MaxGhenis]

Summary

Fixes the framework-level root cause of PolicyEngine/policyengine-us#8780.

SingleAmountTaxScale.calc() returned an int64 0/1 array for boolean single_amount brackets (amount_unit: bool with amount: true/false), not a boolean array. Applying ~ (logical NOT) to that int result performed a bitwise negation (~1 == -2, ~0 == -1, both truthy) instead of a logical one, silently producing wrong values in consuming formulas. The bug was masked whenever the surrounding expression happened to &/| the result with a genuine bool term, so it shipped undetected until a masking term was removed during cleanup in policyengine-us (#8763).

Root cause

In single_amount_tax_scale.py, the amounts array was built as:

guarded_amounts = numpy.array([0] + self.amounts + [0])

Even when self.amounts are Python booleans ([True, False]), prepending/appending the Python int 0 sentinels forces numpy to infer int64 for the whole array, dropping the boolean dtype. ~int64 is then a bitwise negation.

Fix

Build the amounts array from the bracket amounts first, so numpy infers their natural dtype (bool for boolean brackets, int/float otherwise), then pad the two out-of-range sentinel cells with a zero of that same dtype (False for bool, 0 for numeric):

if len(self.amounts) == 0:
    guarded_amounts = numpy.array([0, 0])
else:
    amounts = numpy.asarray(self.amounts)
    sentinel = numpy.zeros(1, dtype=amounts.dtype)
    guarded_amounts = numpy.concatenate([sentinel, amounts, sentinel])

Why this is low-risk (no ripple into numeric consumers)

• Numeric int/float brackets are untouched — same dtype, same values (verified: int stays int64, float stays float64).
• Empty scales still return int64 [0] (historical behavior preserved via the len == 0 guard).
• The only operator whose semantics differ between an int 0/1 array and a bool array is ~ — which is exactly the bug being fixed. Every other combinator used by consuming formulas (&, |, where(), ==, *, +) behaves identically on bool and int 0/1, because numpy auto-promotes bool in arithmetic. So switching boolean brackets from int64 to bool is strictly a correctness improvement with no numeric regression.

With this fix, the per-site .astype(bool) workarounds already merged in policyengine-us (#8919 and the aca_required_contribution_percentage fix) become redundant but remain harmless; they are intentionally left in place.

Tests

Added to tests/core/tax_scales/test_single_amount_tax_scale.py:

• test_calc_preserves_boolean_dtype — bool brackets return a bool array.
• test_calc_boolean_result_supports_logical_not — ~ on the result is logical NOT ([False, True]), the exact regression guarded.
• test_calc_preserves_integer_dtype / test_calc_preserves_float_dtype — numeric dtypes and values unchanged.
• test_calc_out_of_range_bool_returns_false_sentinel — out-of-range sentinel is False, not int 0.
• test_calc_empty_scale_stays_int — empty scale still returns int64 [0].
• test_calc_boolean_dtype_end_to_end_via_parameter_scale — full ParameterScale → get_at_instant → SingleAmountTaxScale path with an amount_unit: bool scale (mirroring how policyengine-us declares these), asserting bool dtype and correct ~.

Checks run (Python 3.13, editable install in a clean worktree):

• tests/core/tax_scales/test_single_amount_tax_scale.py: 12 passed (5 pre-existing + 7 new).
• tests/core/tax_scales/ + tests/core/parameters/: 89 passed.
• Full suite tests/: 615 passed, 4 skipped, 1 xfailed — no regressions.
• ruff format and ruff check clean on both changed files.

Changelog fragment: changelog.d/bool-bracket-dtype.fixed.md.

[MaxGhenis]

Lead review: root cause exactly identified — the Python int sentinels in calc() coerced boolean bracket amounts to int64, making ~ bitwise (both ~0 and ~1 truthy) instead of logical. Fix preserves amounts dtype with same-dtype sentinel padding and keeps the empty-scale int64 fallback. Numeric brackets are byte-identical before/after; verified via full core suite (615 passed) and a 1,004-test pe-us consumer sweep with this branch installed. The three previously-fixed pe-us .astype(bool) sites stay as harmless belt-and-suspenders.