feat: add YAML policy-test harness

[vahid-ahmadi]

Summary

Closes #47 in part — ships the runner and a starter set of YAML cases. Stacked on #53 (parity harness) → #52 (from_situation).

YAML cases mirror the format used by policyengine_uk/tests/policy/, so the existing 196 Python YAML tests can be ported into this repo one at a time without touching the runner.

# tests/policy/income_tax.yaml
- name: Spans basic and higher rates
  period: 2025
  absolute_error_margin: 1
  input:
    employment_income: 50_000
  output:
    baseline_income_tax: 7486

Run them:

$ python -m policyengine_uk_compiled.yaml_tests tests/policy
  PASS  Zero income — no tax [income_tax.yaml]
  PASS  Spans basic and higher rates [income_tax.yaml]
  PASS  Two children, modest income — full Child Benefit [child_benefit.yaml]
  ...
  11 passed, 0 failed of 11

…or pick them up automatically as parametrized pytest cases:

$ pytest interfaces/python/tests/test_yaml_policy_cases.py
test_yaml_policy_case[Zero income — no tax] PASSED
test_yaml_policy_case[Spans basic and higher rates] PASSED
...

What's included

Runner (interfaces/python/policyengine_uk_compiled/yaml_tests.py):

• Two input forms: flat single-person shorthand (employment_income: 50000 → auto-builds a one-person household) and full situation (people: / benunits: / households: matching from_situation).
• Variable name routing on the output side: each expected variable is auto-located in persons, benunits, or households microdata and summed across rows.
• Absolute + relative error margins.
• Boolean comparison treated as exact (so a £1 numeric tolerance doesn't make False ≈ True).
• CLI: python -m policyengine_uk_compiled.yaml_tests <path> (file or dir).

Initial YAML cases (tests/policy/, 11 cases):

• income_tax.yaml: zero income, below PA, basic-rate band, basic↔higher span, fully-tapered PA at £150k
• national_insurance.yaml: below PT, main band, partial-higher tier, well above UEL
• child_benefit.yaml: two children with full CB, no children = £0

Pytest integration (test_yaml_policy_cases.py): auto-discovers every YAML case and runs each as a parametrized test, skipping cleanly when the Rust binary isn't built.

Runner unit tests (21 cases in test_yaml_test_runner.py): cover input routing, tolerance edge cases (incl. boolean comparison), YAML parsing, and validation errors.

pyyaml added to runtime dependencies in pyproject.toml so the runner is usable from the installed package.

Verified locally

• cargo test: 135 passing
• pytest interfaces/python/tests: 69 passing (37 from feat: add Simulation.from_situation for situation-JSON input #52/feat: add Python ↔ Rust parity harness #53 + 21 runner unit tests + 11 YAML cases)
• python -m policyengine_uk_compiled.yaml_tests tests/policy: 11/11 passing
• Boolean tolerance regression caught & fixed during development.

Stacking

vahid/yaml-test-harness ← vahid/parity-harness (#53) ← vahid/from-situation (#52). Once #52 and #53 merge, GitHub will auto-rebase this onto main.

Test plan

• CI passes (cargo + pytest + parity smoke + YAML cases)
• Local: python -m policyengine_uk_compiled.yaml_tests tests/policy prints 11 passes
• A new YAML test under tests/policy/ is automatically picked up by pytest without code changes

🤖 Generated with Claude Code

[vahid-ahmadi]

Rebased onto main and made self-contained — ready to merge.

This was previously stacked on #53/#52 and imported _situation_to_dataframes from the (contested) from_situation PR. The situation→DataFrame conversion is now inlined as a private helper inside the harness, so:

• it no longer depends on feat: add Simulation.from_situation for situation-JSON input #52, and
• it does not add a new public input schema (addressing @nikhilwoodruff's single-entry-point concern).

engine.py is byte-identical to main. Verified: the wrapper imports cleanly, the 21 runner unit tests pass, and all 11 committed YAML policy cases run through the real release binary and pass numerically. Assertion logic is sound (booleans compared exactly, relative margin skipped at expected 0, clear "column not found" failures). Mergeable against main.