Add partner YAML tests and a dedicated CI runner

[hua7450]

Summary

Adds a dedicated, easily-debuggable test layer that fails CI when a PR would change calculation results for any of PolicyEngine's household API partners. Three batches under policyengine_us/tests/policy/baseline/partners/, plus a standalone CI runner so partner-affecting regressions show up as their own named check.

1. Customer fixture mirrors (original scope)

Mirrors policyengine-household-api customer fixtures (Amplifi, Impactica, MyFriendBen) — one folder per partner, one file per year.

2. analytics_coverage/signatures/ — production request coverage (81 cases)

One case per unique input-variable shape partners actually sent through the household API (sourced from unique_signatures.csv — 180 unique signatures collapse to ~83 unique input shapes after deduplicating household-size variants). Grouped by state into 8 files. Each retained case is annotated with how many production variants it consolidates (e.g. signature_3 (54 reqs, +12 dupes)).

3. analytics_coverage/edge_cases/ — boundary tests (642 cases)

Two-level structure mirroring PolicyEngine's own gov/ layout:

analytics_coverage/
├── signatures/      ← per-state signature cases (8 files, 81 cases)
└── edge_cases/
    ├── federal/
    │   ├── tax_credits/   (eitc, ctc, cdcc)
    │   ├── nutrition/     (snap, wic, csfp, school_meals)
    │   ├── healthcare/    (medicaid, chip, aca_ptc, msp)
    │   ├── childcare/     (ccdf, head_start, early_head_start)
    │   ├── cash/          (ssi, tanf)
    │   ├── housing/       (housing_assistance)
    │   ├── utility/       (lifeline)
    │   └── composition/   (cross-program: tax_credits, cash_health)
    └── state/
        ├── ca/  (18 programs: ala_general_assistance, calworks_*, eitc, yctc, tanf, …)
        ├── co/  (7)
        ├── il/  (11)
        ├── la/  (1: general_relief)
        ├── ma/  (8)
        ├── nc/  (2: scca, tanf)
        ├── tx/  (5: ccs, dart, fpp, harris_rides, tanf)
        └── wa/  (3: working_families_tax_credit, tanf, tax_credits_composition)

For each program, cases target the actual binding thresholds (income at FPL boundaries, age cutoffs, asset limits, immigration status, household composition).

State-varying programs (9-state coverage each):

• SNAP (87 cases): per-state binding gross income limit (federal 130% FPL vs state BBCE), age 60 elderly boundary with assets between $3k/$4.5k, disability flip, immigration LPR/REFUGEE/UNDOCUMENTED (captures 2025-07 rule change), plus 6 size/composition variants in federal.yaml.
• Medicaid / CHIP / ACA PTC / SSI / CCDF / housing / school_meals / Head Start / Early Head Start: state-varying eligibility (Medicaid expansion vs not, state copay schedules, FMR, regional rates) with state-invariant cases (federal MAGI categories, federal age cutoffs) deduplicated to federal.yaml.

Federal-only programs: eitc, cdcc, csfp, lifeline, wic, msp, composition/tax_credits — single federal.yaml, verified to produce identical outputs across all states.

State-specific programs: state TANF (8 states), CA tax credits + CalWORKs + general assistance + utility programs (16), CO state credits + supplements (5), IL state programs (9), MA state programs (6), NC SCCA, TX childcare + transit (4), WA WFTC.

What this catches

Verified by deliberately introducing each change locally and observing which suites fail:

Change	Caught by
Variable removed / renamed without alias	all 3 layers (simulation build fails)
Formula refactor that changes outputs	edge_cases (sometimes signatures + fixtures, depends on household composition)
Enum member removal	all 3 layers
value_type, definition_period, or entity change	all 3 layers
Transitive removal (a non-partner variable that partner outputs depend on)	all 3 layers via dependency graph

Known gaps

• Parameter-value changes that don't bind on an existing test threshold can pass undetected. Edge cases today bind on eligibility boundaries (income at FPL cutoffs, age 60/65, $3k/$4.5k assets). They don't yet exercise:
  • Benefit floors (e.g., TX TANF minimum_grant)
  • Allotment caps when the cap doesn't bind in tested households
  • Intermediate deduction amounts (only top-line outputs are checked)
• Follow-up: add "binds on parameter X" cases for each modified parameter type.

Implementation notes

• Snapshot path matches CI: outputs generated via SimulationBuilder.build_from_dict with set_default_period(year) — the same path the YAML test runner uses.
• absolute_error_margin: 0.1: tight enough that boolean eligibility variables (encoded as 0/1) can't silently false-positive across true / false, but loose enough to absorb float32 rounding noise.
• county_str enum fix: prior versions had county_str: "DALLAS COUNTY" (human-readable). PolicyEngine silently ignored these because county_str has value_type=str — the actual county enum defaulted to ALAMEDA_COUNTY_CA regardless. Fixed: all county_str inputs converted to valid enum names (DALLAS_COUNTY_TX, ARAPAHOE_COUNTY_CO, etc.), and outputs recomputed so CalFresh, ACA PTC, and other county-dependent outputs match the actual county.
• Three-stage redundancy trim: peak 1,642 → 866 (drop pure-federal state duplicates) → 642 edge cases + 81 signature cases. Cumulative ~60% reduction without losing meaningful coverage.
• State-aware edge case thresholds: SNAP binding gross income limit pulls each state's BBCE multiplier dynamically (gov.hhs.tanf.non_cash.income_limit.gross[STATE]); state TANF income boundary found via per-state binary search.

CI integration

• New dedicated job: Household API Partners (extracted from the Full Suite Baseline matrix into a standalone job in both pr.yaml and push.yaml).
• Added to the Publish job's needs: in push.yaml so partner-test failures block release.

Test plan

• All 81 analytics_coverage signature cases pass locally
• All 642 edge_cases pass locally (4-way parallel run: 7:35 wall time)
• 7 customer fixture YAMLs pass locally via make test-yaml-no-structural-other-partners
• make format clean
• Confirmed standalone Household API Partners runner picks up analytics_coverage/ and edge_cases/ under the new layout
• CI passes on this PR

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 78.72%. Comparing base (8cb60e7) to head (eee0f91).
⚠️ Report is 19 commits behind head on main.

Additional details and impacted files

@@             Coverage Diff              @@
##              main    #8273       +/-   ##
============================================
- Coverage   100.00%   78.72%   -21.28%     
============================================
  Files            3     4723     +4720     
  Lines           63    68774    +68711     
  Branches         0      340      +340     
============================================
+ Hits            63    54142    +54079     
- Misses           0    14554    +14554     
- Partials         0       78       +78     

Flag	Coverage Δ
unittests	78.72% <ø> (-21.28%)	⬇️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.