2015-2016 households with a member 65+ fail: Medicare IRMAA reads 2013-2014 AGI, where the 401(k) limit parameter does not exist #9613

Summary

Any 2015 or 2016 household with a member aged 65 or older fails to compute income_tax or household_net_income.

medicare_irmaa_magi_two_years_prior computes adjusted_gross_income at period.offset(-2, "year"), so it asks for AGI in 2013 or 2014. backdate_parameters(first_instant="2015-01-01") does not reach those years, and the AGI chain needs gov.irs.gross_income.retirement_contributions.limit[401k] through elective_deferral_limit, so the calculation raises.

Reproduction

main at 9e5852a:

from policyengine_us import Simulation
for y in (2015, 2016, 2017):
    sim = Simulation(situation={
        "people": {"p": {"age": {y: 66}, "employment_income": {y: 30000}}},
        "households": {"h": {"members": ["p"], "state_code": {y: "TX"}}},
    })
    print(y, sim.calculate("income_tax", y))

• 2015: ParameterNotFoundError: The parameter 'gov.irs.gross_income.retirement_contributions.limit[401k]' was not found in the 2013-01-01 tax and benefit system.
• 2016: the same error at 2014-01-01.
• 2017: OK.

Call chain:

medical_expense_health_insurance_premiums
  -> medicare_part_b_premium
  -> gross_medicare_part_b_premium
  -> medicare_irmaa_magi_two_years_prior
  -> adjusted_gross_income (2013/2014)
  -> irs_employment_income
  -> traditional_401k_contributions
  -> elective_deferral_limit

The premium is computed even when medicare_enrolled is false, because medical_expense_health_insurance_premiums multiplies the premium by enrollment rather than gating the computation.

Possible fixes

• Stop the lag at the model's first year: when period.offset(-2) is before 2015, use the earliest supported year, or return 0.
• Alternatively, backdate the parameters on the AGI path to 2013.

Either way, add a 2015 test with a 65+ person.

Found while fixing Arkansas 2014-2020 income tax in #9611 and #9612. #9612's tests supply itemized_medical_expenses directly to avoid this.