Live staging household parity checks fail: a US variable on the wrong entity, and a 90 s timeout on /calculate-full #4287

What fails

The live staging household checks added in #4274 failed in the release run on cd513dacc (push.yml run 43075278725, job "Run live simulations against the staging candidate", step "Run live staging tests, one pytest worker per test"): 2 failed, 7 passed. Staging promotion depends on that job, so no release of the Public API reaches production until it passes.

1. test_a_v2_household_calculation_matches_api_v1[us-single-california]

The Public API's /us/calculate refuses the check's household with 400:

Household variable income_tax belongs on tax_units, not people, at household.people.adult.income_tax.

The household in tests/integration/test_live_v2_household_parity.py puts income_tax on the person. policyengine_us defines income_tax with entity = TaxUnit, and /calculate has refused a variable on the wrong entity since 3a59b4d14 (May 2026). The simulation service's household function has the same check (household_variable_wrong_entity). The household is wrong, not either calculation. The UK households put income_tax on the person correctly, since policyengine_uk defines it there.

2. test_a_v2_household_calculation_matches_api_v1[us-single-reform-added-variables]

httpx.ReadTimeout on the check's POST /us/calculate-full, after its v2 calculation had finished. The check's client allows 90 s per read. The candidate's request log shows the same request, received at 18:44:19.6 UTC, answered 200 after 223.8 s. The client gave up at 18:45:49.9, 90.3 s in.

That time is normal for /calculate-full, which calculates every variable the household leaves out. Production's /calculate-full requests from the past week took 50 s, 84 s, 248 s and 296 s, and some ended with Cloud Run's 504 at its 300 s request timeout. That timeout is CLOUD_RUN_TIMEOUT in .github/scripts/cloud_run_env.sh, the same for staging and production.

Fix

• Put the US household's income_tax on its tax unit.
• Give the parity check's calls to the Public API's own /calculate and /calculate-full a read timeout longer than the candidate's 300 s request timeout. The check then gets the candidate's own answer.
• Add unit tests that would have caught both failures before a release:
  • each parity household passes the Public API's real calculation routes, input checks and situation parsing included, with the installed country models;
  • the check's request waits longer than CLOUD_RUN_TIMEOUT.