[anth-volk]

Fixes #941

Same-Repository Draft PR Check

• This PR is a draft.
• This PR is from a branch in PolicyEngine/policyengine-sim-api, not a personal fork.

Summary

Seventh of nine stacked PRs; it builds on #938. It splits the executor's simulation_runtime.py and its output builder tests without changing behaviour, and removes both from the file-length baseline (count 4 to 2). Scripts compared every top-level statement of each old file with the new files: all appear exactly once and unchanged apart from the renames below and one logger line, in their original order.

simulation_runtime.py (1,111 lines)

Module	Lines	Holds
gcp_credentials.py	119	writing the worker's Google Cloud credentials
request_policy.py	133	the request's year and its policy periods
region_resolution.py	283	RegionResolution, resolving a request's region or region group, and checking a local-area breakdown request
simulation_runtime.py	651	dataset selection and loading, simulation building, local-area breakdowns, the run, and re-exports

• Why simulation_runtime.py stays at 651 lines, above the 400 the split aims for. Tests across the executor (the runtime, precompute, Stage 12, job-progress and segmented-national tests) replace _country_module, _load_dataset, _build_simulation, _resolve_dataset_selection, resolve_data_folder, SimulationOutputBuilder and others by name on simulation_runtime. A replacement only reaches code that looks the name up in that module, and _load_dataset and _build_simulation themselves call _country_module and resolve_data_folder, so moving any of these apart would silently stop some replacements applying. They stay together; the code moved out reads none of them.
• Logging: gcp_credentials.py names the policyengine_simulation_executor.simulation_runtime logger explicitly, so the credential messages keep their logger; this is the one statement that isn't verbatim.
• Import order: the moved modules import nothing the old module didn't, so POLICYENGINE_SKIP_COUNTRY_IMPORTS is still set before any policyengine import.

Renames

Private names another module now uses are public, each with a docstring where it had none: _parse_year to parse_year, _normalise_policy to normalise_policy, _compile_interval_policy to compile_interval_policy, _runtime_policy to wrapper_policy (its docstring says it converts a request's policy into the form the wrapper receives; _build_simulation already has a local variable named runtime_policy), _resolve_region to resolve_region, _resolve_dataset_reference to resolve_dataset_reference, and _require_local_area_breakdown to require_local_area_breakdown.

simulation_runtime.py re-exports them. The modules that imported the private names from it now import the public names from it, so tests that replace them on simulation_runtime still reach those modules: src/modal/segmented_national.py, spm.py, precompute_planning.py, stage12_cache/precompute_planning.py, stage12_runtime/aggregation.py, stage12_runtime/simulation.py and stage12_worker_validation.py.

Tests: tests/test_simulation_output_builder.py (3,048 lines)

File	Lines
test_simulation_output_fakes.py (shared fakes, no tests)	290
test_simulation_output_builder.py	444
test_simulation_output_impacts.py	338
test_simulation_output_runtime_runs.py	419
test_simulation_output_dataset_loading.py	242
test_simulation_output_dataset_selection.py	354
test_simulation_output_uk_local_areas.py	802
test_simulation_output_region_resolution.py	257

• Every file is named test_simulation_output_*.py, the pattern sim.microdata.builder-tests-use-weighted-frames checks, so the entity frames these tests and fakes build are still checked; the fakes module's docstring says so. test_simulation_output_uk_local_areas.py keeps the routed-UK tests together with the fakes only they use.
• The shared fakes are public: _FakeOutputDataset, _FakeSimulation, _TrackedSpan, _TrackingRuntime, _with_observability_timings, _macro_baseline_reform, _congressional_district_output, _constituency_impact_record, _local_authority_impact_record, _stub_policyengine_output_calls and _build_schema_output drop their underscore, and _simulation_output_builder is make_output_builder.
• All 1,895 executor tests are collected before and after, with the same names.
• Patch targets (a script listed, for each replacement, the modules whose code reads the name):
  • the region-group tests (tests/test_region_group_resolution.py) and the SPM boundary helper (tests/test_spm_runtime_native.py) replace resolve_dataset_reference in region_resolution, where regions are resolved;
  • the identity fixture in tests/test_precompute_planning.py replaces get_country_release_bundle in both simulation_runtime and region_resolution, which both read it, as the one replacement did before;
  • every other replacement still targets simulation_runtime, under the new name where the helper was renamed (resolve_region in seven tests and helpers);
  • tests that import or read a private helper that moved now take it from its new module: _normalize_credentials_blob (gcp_credentials), _parse_policy_period (request_policy), _normalise_region_code and _resolve_region_group (region_resolution).

Environment variables

No change: setup_gcp_credentials reads the same GOOGLE_APPLICATION_CREDENTIALS_JSON and GOOGLE_APPLICATION_CREDENTIALS variables in its new module, and no workflow, image or deployment file changes.

Testing

• Executor, full suite: 1,873 passed, 22 skipped, as before; the weighted-frame guard finds nothing in the twelve files it now scans.
• Entrypoint: 742 passed. Gateway: 268 passed.
• Pyright for the executor: no errors.
• Ruff 0.15.22 format and lint: clean.
• python scripts/run_quality_guards.py and python -m scripts.rules guard --base-ref origin/file-length-split-precompute: passed; python -m scripts.rules render leaves no diff.

Only the lowest PR in the stack runs full CI, so this PR's CI complete check fails by design until the PRs below it merge.

🤖 Generated with Claude Code