Tighten maintenance loan proxy logic

[MaxGhenis]

Summary

• split maintenance loan eligibility into explicit helper variables for England-system scope and higher-education evidence
• tighten the sponsor, living-arrangement, and assessed-household-income proxies using a fast two-candidate adult proxy plus sponsor benefit-unit income
• add focused policy and Python regression coverage for higher-education evidence precedence and sponsor-income edge cases

Validation

• uv run policyengine-core test policyengine_uk/tests/policy/baseline/gov/dfe/maintenance_loans/maintenance_loan.yaml
• uv run pytest -q policyengine_uk/tests/test_maintenance_loan_proxies.py
• full 2025 maintenance-loan chain on enhanced_frs_2023_24.h5: candidate benunit income 0.02s, sponsor detection 0.05s, sponsor income 0.01s, household income 0.01s, award 0.02s

Aggregate Impact

Measured on enhanced_frs_2023_24.h5 at 2025, comparing merged #1586 against this branch on current weights.

• maintenance_loan mean: 143.9555 -> 162.5093 (+18.5539, +12.9%)
• maintenance_loan recipient mean: 8057.7370 -> 9096.2669 (+1038.5298, +12.9%)
• maintenance_loan recipient rate: unchanged at 1.78655%
• household_tax mean: unchanged at 24904.1915
• hbai_household_net_income mean: unchanged at 44999.9928
• relative poverty AHC: unchanged at 21.4478%
• relative poverty BHC: unchanged at 17.2909%

Caveat

• maintenance_loan_in_england_system still proxies Student Finance England membership from current household country, so cross-border domicile cases remain approximate.

[MaxGhenis]

2025 aggregate impact on current weights versus merged :\n\n- mean: (, )\n- recipient mean: (, )\n- recipient rate: unchanged at \n- mean: unchanged at \n- mean: unchanged at \n- relative poverty AHC: unchanged at \n- relative poverty BHC: unchanged at \n\nLocal validation on the final rebased head:\n- WARNING:policyengine_core.scripts:Several country packages detected : policyengine_uk, policyengine_core.country_template. Using policyengine_uk by default. To use another package, please use the --country-package option.
============================= test session starts ==============================
platform darwin -- Python 3.13.9, pytest-8.4.1, pluggy-1.6.0
rootdir: /Users/maxghenis/worktrees/policyengine-uk-fix-560-followup
configfile: pyproject.toml
collected 21 items

policyengine_uk/tests/policy/baseline/gov/dfe/maintenance_loans/maintenance_loan.yaml .....................

============================== 21 passed in 0.20s ==============================\n- ..... [100%]
=============================== warnings summary ===============================
.venv/lib/python3.13/site-packages/pydantic/_internal/_config.py:323
/Users/maxghenis/worktrees/policyengine-uk-fix-560-followup/.venv/lib/python3.13/site-packages/pydantic/_internal/_config.py:323: PydanticDeprecatedSince20: Support for class-based config is deprecated, use ConfigDict instead. Deprecated in Pydantic V2.0 to be removed in V3.0. See Pydantic V2 Migration Guide at https://errors.pydantic.dev/2.11/migration/
warnings.warn(DEPRECATION_MESSAGE, DeprecationWarning)

-- Docs: https://docs.pytest.org/en/stable/how-to/capture-warnings.html
5 passed, 1 warning in 2.20s\n\nThe remaining known approximation is that still proxies Student Finance England membership from current household country, so cross-border domicile cases remain approximate.

[MaxGhenis]

Dataset-backed aggregate check on enhanced_frs_2023_24.h5 at 2025, comparing merged #1586 against this follow-up branch on the same current weights:

• maintenance_loan mean: 143.96 -> 162.51 (+18.55, +12.9%)
• maintenance_loan recipient rate: 1.78655% -> 1.78655% (no change)
• maintenance_loan recipient mean: £8,057.74 -> £9,096.27 (+£1,038.53, +12.9%)

The broader household aggregates I checked were unchanged to displayed precision:

• household_tax mean: no change
• hbai_household_net_income mean: no change
• relative poverty BHC: no change
• relative poverty AHC: no change

So the follow-up is tightening the maintenance-loan proxy and raising award amounts for the same modeled recipient pool, without moving the broader tax/poverty aggregates on current weights.