Document month-keyed YAML test outputs

[hua7450]

Summary

The testing guidance says the YAML runner supports only YYYY and YYYY-01, and concludes a mid-year month cannot be tested. The restriction is real but applies to the case period:; keying the output by month inside a whole-year period asserts any month.

- name: Case 41, PA waiver starts in September 2024.
  period: 2024
  input:
    state_code: PA
    county_fips: "42005"
  output:
    is_in_snap_abawd_waived_area:
      2024-08: false
      2024-09: true

Fixes #80

Why this matters

The old advice does not just omit a technique, it recommends a worse test. references/tests.md said to "test with the first month of a year where it's fully active"; test-creator said to use the next January after the effective date. Both yield a case that passes whether or not the mid-year boundary is modeled — a silently vacuous test rather than a failing one. Asserting the months on either side of a change is what proves the boundary sits where you think it does.

It also pushed authors toward hand-built Simulation objects for something the runner handles natively. That happened in PolicyEngine/policyengine-us#9243: a Python test file was written for an April 1 boundary, and the PR description asserted on the strength of this guidance that the boundary "cannot be expressed there." Both have since been corrected.

Mid-year boundaries are common in benefit rules — state BBCE standards re-basing off the federal October cycle, ABAWD waiver windows, mid-year rate changes — so this is not a rare corner.

Changes

• references/tests.md — scopes the restriction to period:, adds an "Asserting a specific month" section with the pattern, when to prefer it, and the month-naming failure message. Keeps the existing (correct) statement about which period: values fail.
• test-creator — same correction at the point of use, and clarifies that month keys under output: are not what its period: self-check targets, so the agent does not "fix" valid tests.

The mid-year January fallback is retained, rescoped to cases that are not asserting the boundary.

Verification

Per authoring rule 1, confirmed by execution rather than memory. On tanf_non_cash_gross_income_limit in policyengine-us: 14 cases pass with month-keyed outputs, and deliberately corrupting one month fails with

tanf_non_cash_gross_income_limit@2026-04: [2660.] differs from 9999.0 with an absolute margin > 0.01

confirming the assertions evaluate rather than being parsed and ignored — the failure mode that would otherwise make this change worthless.

Test plan

• uv run pytest --ignore=tests/test_skill_examples.py — 99 passed
• python3 scripts/build_claude_wrapper.py — builds clean
• CI passes