[anth-volk]

Fixes #4025

What and why

Since its move to .py 4.x, the simulation API has returned decile.relative and the UK wealth_decile.relative as percentages (multiplied by 100) instead of fractions. Production economy results have carried these values since 2026-06-29, when the Public API started requesting .py 4.18.3. PolicyEngine/policyengine-sim-api#827 tracks the bug, and PolicyEngine/policyengine-sim-api#904 fixes it by dividing by 100 in build_decile_output.

The Public API caches economy results in Redis for 30 days (REFORM_IMPACT_TTL_SECONDS, with jitter). A lookup matches on the scope (country, policy pair, region, dataset and period), the cache version (ReformImpactIdentity.api_version, from get_economy_impact_cache_version) and the options hash. The cache version is a digest of the country package, .py and policyengine-core versions plus RUNTIME_CACHE_SCHEMA_VERSIONS["economy_impact"]. The options hash covers the request options, the dataset, the model and .py versions, and the simulation app name. None of these change when only the simulation API's code changes, so after #904 is deployed, results already cached for .py 6.7.5 would keep being served with multiplied-by-100 values for up to 30 days.

This PR raises RUNTIME_CACHE_SCHEMA_VERSIONS["economy_impact"] from 3 to 4. The cache version changes from e3… to e4…, so lookups read a different Redis scope index and don't find results stored under version 3. The next request for each economy result recomputes it.

Budget-window cache

BudgetWindowCache.build_key takes its api_version from the same get_economy_impact_cache_version call (through _build_economic_impact_setup_options), so this change also gives every budget-window result a new key, and budget-window results are recomputed once. Budget-window results hold only yearly budget totals (tax revenue, benefit spending and budgetary impact), not decile output, so BUDGET_WINDOW_CACHE_SCHEMA_VERSION doesn't need its own change.

When to merge

Only after PolicyEngine/policyengine-sim-api#904 is merged and deployed to the simulation app the Public API uses, which is policyengine-simulation-py6-7-5 today. Merging earlier would recompute results that are still multiplied by 100 and cache them under the new version for another 30 days.

Interaction with the .py 6.10.0 update (#4024)

#4024 updates .py to 6.10.0, which changes both the cache version digest and the options hash, so results cached for 6.7.5 stop being reused once it ships, with or without this PR. That only helps if the simulation app for .py 6.10.0 also includes #904's fix. Without it, that app computes and caches new multiplied-by-100 results under the new key.

What this doesn't fix

Report outputs already saved to v1 report records keep the multiplied-by-100 values, because they are stored in the database and are not read back from the economy cache. Correcting them is a follow-up. RUNTIME_CACHE_SCHEMA_VERSIONS["report_output"] is unchanged here.

Tests

• tests/unit/test_constants.py::TestEconomyImpactCacheVersion pins RUNTIME_CACHE_SCHEMA_VERSIONS["economy_impact"] at 4, and checks that the US and UK cache versions start with e4 and still fit the 10-character api_version column.
• tests/unit/services/test_reform_impacts_service.py::test_impact_cached_under_previous_economy_schema_version_is_not_matched stores a completed impact under the version-3 cache version in the in-memory cache, then checks that the exact and options-hash-prefix lookups with the current cache version return nothing, while a lookup with the old version still finds it.

With the version still at 3, all four new tests failed (assert 3 == 4, 'e30e9091da'.startswith('e4'), and the version-3 impact returned by the current-version lookup). They pass at 4.

Local checks:

• make format, ruff check . and ruff format --check .: clean
• uv run --frozen --extra dev mypy: no issues
• focused tests (constants, runtime cache, reform impacts, budget-window cache, economy service, report output service): passed
• make test: 3285 passed
• python scripts/run_quality_guards.py: passed

Environment variables

None change. This PR changes one Python constant, its tests and a changelog fragment.

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 87.72%. Comparing base (881a009) to head (beadab0).

Additional details and impacted files

@@             Coverage Diff             @@
##           master    #4026       +/-   ##
===========================================
+ Coverage   51.11%   87.72%   +36.60%     
===========================================
  Files         242      258       +16     
  Lines       14441    15864     +1423     
  Branches     2307     2594      +287     
===========================================
+ Hits         7382    13917     +6535     
+ Misses       6436     1280     -5156     
- Partials      623      667       +44     

☔ 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.

[PavelMakarchuk]

Thanks Anthony, this looks right. I approve.

I confirmed the precondition: the prod deploy of #904 (sim-api run 39879445060, commit 1fe19571) redeployed policyengine-simulation-py6-7-5, so new economy results from the Public API now carry fractions. It's safe to merge this now.

The change and the tests look good: version 4 makes the exact and the prefix lookups miss the version-3 entries, and the new tests fail at 3. The one-time recompute of budget-window results is fine, since they hold no decile output.

Agreed that the ×100 values in saved v1 report records need a follow-up. Is there an issue for it yet?