Carer's Allowance earnings-limit property test counts pence in float64, not the model's float32 #2532

policyengine_uk/tests/test_carers_allowance_earnings_limit_properties.py::test_bounds_and_differential can fail at random. Its reg. 5 penny-rounding check counts pence in a different precision from the formula it tests.

What goes wrong

_person() turns each result into a Python float (float64) with float(np.asarray(...)[0]). The test then recomputes the reg. 5 rounding, np.floor(earnings / WEEKS * 100 + 1e-6) / 100 > LIMIT, in float64. The model stores carers_allowance_earnings as float32, and is_gainfully_employed_for_carers_allowance floors the weekly pence in float32. Where the formula's float32 arithmetic rounds a value across a penny boundary, the two disagree.

Case

Pay of £10,608.52 in 2026, with no profit and no pension contributions (measured on main at c6715d4):

The assertion bool(result["is_gainfully_employed_for_carers_allowance"]) == over then fails. The test isn't derandomized, so a draw in that narrow band would turn CI red with no code change.

The test should mirror the formula's float32 arithmetic. The float64 recompute applied reg. 5 exactly to the stored value, which is £204.00 a week (20,400.9991 pence floored); the formula's float32 rounding gives £204.01, which matches the £10,608.52 entered. Whether the formula should count pence in float32 at all is #2541. #2518 found and fixed the same pattern in test_carers_allowance_pension_contribution_deduction_rate_properties.py (calculate(), over_limit() and the PENNY_OVER example).

Fix

Test only: no model source changes.

• Count the pence on the stored float32 value, as the formula does. Convert with float() only for the bounds and tolerance checks.
• Add an explicit Hypothesis @example at pay 10,608.52, profit 0, pension 0.
• Derandomize the property tests in this file, so CI draws the same examples every run.