[anth-volk]

Fixes #4149

Summary

CI no longer loads the complete installed PolicyEngine.py catalog (about 1.2 million parameter values). Two tests replace the two full-catalog tests, and both run on every run of the cross-database check, with no conditions.

• Removed: tests/integration/test_v2_catalog_installed.py, tests/integration/test_v2_catalog_publication_qualification.py, their steps in catalog-runtime-integration, and the RUN_V2_CATALOG_COMPATIBILITY and RUN_V2_CATALOG_PUBLICATION_QUALIFICATION switches, which nothing else read. "Cross-database integration tests" keeps its single job, catalog-runtime-integration ("MySQL, PostgreSQL, and Redis integration"); pr.yml, push.yml, "CI complete" and .github/release_commit.py are unchanged.
• New step, "Publish selected parameter folders of the installed PolicyEngine.py models" (tests/integration/test_v2_catalog_installed_folders.py, with --durations=10):
  • It loads the real installed US and UK models, as extract_installed_catalog() does, and gives extract_catalog a reduced view of each. The view keeps the real model identity, package and release manifest, only the national region, no variables, and only the parameters and parameter nodes under the folders below.
  • It publishes the result to the disposable PostgreSQL twice; the second publication compares every stored row with the catalog in both directions. It then reads every parameter value, parameter, node, variable and region back and compares them with the extracted catalog as strict JSON, which tells 1, 1.0 and true apart.
  • It fails with the folder's name when a listed folder no longer exists in the installed models. A second check fails when the folders together stop holding any of the value shapes listed below.
• Edge-case catalog: tests/fixtures/v2_catalog_value_shapes.py covers the shapes extraction accepts that the installed models don't hold (text, mappings, enums as values, NumPy scalars and others). It is published and read back by test_v2_catalog_publication.py::test_publication_stores_every_value_shape_unchanged. The row comparison helpers both tests use moved to tests/integration/catalog_rows.py.

The selected folders

Model	Folder	Why it's there
US	gov.irs.income.bracket	income tax rates and thresholds by bracket number and by filing status (SINGLE, JOINT, SEPARATE, HEAD_OF_HOUSEHOLD, SURVIVING_SPOUSE); the top bracket's thresholds are infinite
US	gov.irs.payroll.social_security	plain rates and a cap, with a long history of values
US	gov.aca.family_tier_states	booleans by state, some starting in year 1 (.py reads Core's 0000-01-01 as year 1)
US	gov.aca.ineligible_immigration_statuses	a list of enum names
US	gov.irs.credits.premium_tax_credit.phase_out	bracket scales (ending_rate[0].amount, ...) whose upper brackets hold null values
US	gov.states.ne.dhhs.child_care_subsidy.fpg_fraction	values in force for a single day, the only ones in the installed models
UK	gov.hmrc.income_tax.rates.uk	the UK income tax bracket scale: a rate and a threshold per bracket
UK	gov.dwp.state_pension.triple_lock	UK booleans and a null value

Together that is 150 parameters and 2,913 parameter values (US 136 and 2,870, UK 14 and 43). The installed models hold integers, fractions, booleans, lists, infinities and nulls, and no plain text or mapping values. Enum names appear as breakdown keys and inside lists. Every value is open-ended or ends the day before the next one starts.

What this gives up

• The 256 MB limit on publish_catalog's additional memory while it publishes the complete catalog, which test_v2_catalog_publication_qualification.py measured with tracemalloc.
• Publishing the complete catalog before merge. Each release still publishes and validates the complete catalog against staging, then production (seed-v2-database.yml, scripts/publish_v2_metadata_catalog.py), before each deploy job. That compares the full catalog with the stored rows on every release, but inserts the full catalog only when the .py version is new.
• The exact per-bundle catalog counts that test_v2_catalog_installed.py pinned (7,438 variables, 29,367 parameter nodes, 104,673 parameters, 1,196,139 parameter values, 4 datasets, 1,837 regions for the 6.11.1 bundle), and its other checks on the complete catalog: the dataset each US and UK region type defaults to, England's region filter, the US place fallback count, and that every model version is .py's version.

Cost of loading the models

Importing policyengine builds both country models; after that, policyengine.us.model and policyengine.uk.model are attribute reads. The test's first test uses only the loaded models, so --durations reports their loading on its own.

Where	Import of .py, building both models	Extracting the selected folders	Publishing and reading back	Step
Local (macOS, warm)	12.2–14.7 s (a standalone import policyengine took 14.6 s)	0.15 s	0.22 s	13.6–16.4 s for the test file
CI, run 41457077599	running	running	running	running

Before this change, over the last 13 successful pull request runs (2026-10-29 and 2026-10-30), "Cross-database integration tests / MySQL, PostgreSQL, and Redis integration" took 7.2 to 37.6 minutes (median 11.1). Its two full-catalog steps took 5.0 to 32.8 minutes of that; every other step together took 2.1 to 4.8 minutes (median 3.1).

Environment variables and secrets

Variable	Read by	Provider	Wiring
CODECOV_TOKEN	the Codecov upload	GitHub Actions secret, both a repository secret and an organization secret (visibility: all repositories); the repository secret takes precedence	pr.yml (job v2-integration-check) and push.yml (job v2-integration-check) pass it explicitly as secrets: CODECOV_TOKEN: ${{ secrets.CODECOV_TOKEN }}; the reusable workflow declares it as an optional workflow_call secret. A workflow_dispatch run reads the repository secret directly. Unchanged.
V2_MIGRATION_DATABASE_URL, V2_ALEMBIC_DISPOSABLE_TEST	the new test, through tests/integration/conftest.py's disposable_v2_database_url	literals in the job's env:; the URL points at the job's own postgres:17 service	job scope, unchanged
UV_PYTHON, ALEMBIC_DATABASE_URL, DEPLOYMENT_ENVIRONMENT, RUNTIME_CACHE_TEST_URL	the other steps	literals in the job's env:	job scope, unchanged
POLICYENGINE_SKIP_COUNTRY_IMPORTS	import policyengine (country models load unless it is "1")	not set anywhere in the workflow	none; when unset, both models load, which the test needs

No other variable is read on the new test's path. The full-catalog switches RUN_V2_CATALOG_COMPATIBILITY and RUN_V2_CATALOG_PUBLICATION_QUALIFICATION are gone.

Testing

Run individually, by file or node ID (no full suite):

• tests/integration/test_v2_catalog_installed_folders.py against a local postgres:17 (four tests, including a missing-folder failure)
• tests/integration/test_v2_catalog_publication.py::test_publication_stores_every_value_shape_unchanged; it fails when one expected value is changed from 12000.0 to 12000
• tests/unit/test_v2_integration_workflow.py (replaces test_full_catalog_integration_workflow.py): one job, no conditional step, no schedule, no full-catalog test, file or switch, and the new step after the role checks
• tests/unit/v2/test_catalog_value_shapes.py
• tests/unit/test_alembic_workflows.py::test_reusable_v2_integration_check_uses_disposable_databases_and_coverage (one coverage run --branch, seven coverage run -a --branch)
• tests/unit/test_locked_test_environments.py::test_detects_every_known_test_job (EXPECTED_TEST_JOBS back to catalog-runtime-integration only) and test_test_job_selects_the_production_python_version[v2-integration-check.yml:catalog-runtime-integration]
• python scripts/run_quality_guards.py (all guards pass) and python -m scripts.rules render (no changes; rules/registry.toml names neither deleted test)

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 88.14%. Comparing base (c4e3a92) to head (18f4c89).
⚠️ Report is 11 commits behind head on master.

Additional details and impacted files

@@             Coverage Diff             @@
##           master    #4150       +/-   ##
===========================================
+ Coverage   50.84%   88.14%   +37.30%     
===========================================
  Files         270      286       +16     
  Lines       15221    16659     +1438     
  Branches     2443     2733      +290     
===========================================
+ Hits         7739    14684     +6945     
+ Misses       6841     1300     -5541     
- Partials      641      675       +34     

☔ View full report in Codecov by Harness.
📢 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.