[anth-volk]

Fixes #870

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

Fourth in the Pyright stack, on top of #869. Fixes all 54 Pyright errors in the executor's simulation_runtime.py without changing what a run does. The executor's baseline goes from 126 entries to 72.

Commit	Errors fixed	Baseline after
Open the simulation runtime's spans with the typed span scope	49	77
Type the simulation runtime's policy, region and dataset reads	5	72

No scopes.operation helper was needed: this module opens only spans.

What changed

• Spans. The nine runtime.span(...) scopes become scopes.span(runtime, ...), including the credential-setup span chosen by a conditional expression, for consistency (Pyright didn't flag that one). That removes the 47 possibly-unbound errors.
• _build_simulation's return type. With baseline and reform bound, Pyright reported that a Simulation can't be passed to _report_calculation_start, which takes a StageReportingSimulation (the 2 errors were already baselined, worded Unbound | Simulation). _build_simulation returns one in both branches (ArtifactBaselineSimulation subclasses it), and _report_calculation_start relies on that by calling on_first_ensure, so the return annotation now says StageReportingSimulation. The unused Simulation import goes.
• Compiled reform. compile_reform returns None only for an empty policy, which _compile_interval_policy has already returned unchanged. A None result now returns the policy too (unreachable), as policyengine's own _reform_dict_to_parameter_values does.
• UK region type. prefix not in {"constituency", "local_authority"} didn't narrow prefix, because a set display is typed set[str]. A tuple keeps the literal types, so prefix narrows to the two values Region.region_type accepts.
• Input dataset entity data. The UK weight-matrix check counts households from dataset.data.entity_data, but policyengine types Dataset.data as BaseModel | None. It now reads entity_data with getattr(..., None) and requires a dict, raising TypeError with the dataset's name when none is loaded; before, that case raised AttributeError. A loaded dataset counts the same households. Narrowing on the value, rather than isinstance(..., UKYearData), keeps the existing tests' SimpleNamespace datasets valid. It also needs no new import into a module that keeps policyengine out of its module-level imports.
• Compiled policy variable. _build_simulation reassigned its policy: dict | None parameter to _compile_interval_policy's result, which may be a Policy. The result now has its own name, runtime_policy, used in the three places that read the reassigned name.

The microdata weight checks (sim.microdata.*) recognise weighted frames by expressions such as <simulation>.output_dataset.data.<entity>. The dataset read changed here is the input dataset (dataset.data), which they don't track, before and after; test_microdata_weight_boundaries.py passes.

Commits (red, then green)

1. Red: with the 49 entries removed from the baseline, the check on the code before the fix reported 49 errors it doesn't list. Green: passes with 77; executor tests pass.
2. Red: with the module's last 5 entries removed, the check reported them. Green: passes with 72; executor tests pass, including test_uk_weight_matrix_guard.py.

Suppressions added

None.

Real bugs found

None.

Self-review

review-code on this PR's diff, run by a separate agent, found no bugs and no material code-quality issues. It confirmed:

• the 54 removed entries are exactly this module's;
• each commit's Pyright output equals its baseline;
• each change preserves what a run does.

It noted that the new TypeError for an input dataset with no entity data is untested, and that a dataset loaded from a file can't reach it. #873 switches that check to the shared EntityFrames protocol.

Environment variables

No CI, image or deployment configuration changes, and no new environment-variable reads. simulation_runtime.py runs in the Modal simulation workers; it now imports policyengine_simulation_observability.scopes (#864), part of the observability library the images already install.

Rebase onto #879 (2026-10-22)

Rebased onto the rebuilt #869 after #879 merged. Where this section and the text above differ, this section is current.

What conflicted. simulation_runtime.py, in the first commit. #879 had already opened the run's eight request, dataset, policy and simulation-build spans in _run_simulation_impl_core with scopes.span, and annotated _build_simulation as returning StageReportingSimulation. That removed the same 49 errors this commit removed.

How it was resolved. #879's code is kept, including its baked per-state dataset lookup inside the dataset-load span. The first commit now only opens the credential-setup span in setup_gcp_credentials with scopes.span, and adds _build_simulation's docstring. Neither changes the error count, so its baseline stays at 77. The second commit applied cleanly and removes the module's last 5 errors.

Commit	Errors fixed	Baseline after
Open the simulation runtime's spans with the typed span scope	0 (#879 fixed the 49)	77
Type the simulation runtime's policy, region and dataset reads	5	72

Baseline at this PR's head: 72 entries, the same entries as before the rebase.

Checks after the rebase. Pyright passes with 72 at both commits. The precompute, precompute app, output builder, state datasets, artifact keys, image setup, bundle image, GCP credentials, runtime policy, UK weight-matrix and annotation tests pass (314). Ruff 0.15.22 format and check are clean. The quality guards and the rules guard against #869 pass. render leaves no changes. No new noqa.

Rebase onto main after #904 and #881 (2026-10-26)

Rebased onto main 34209283, then onto 1f985a73 after #862 merged, which changed nothing in this PR. Where this section and the text above differ, this section is current.

What conflicted. simulation_runtime.py, rules/baselines/pyright-executor.txt, rules/registry.toml and docs/generated/rules.md. Both conflicting blocks in simulation_runtime.py were in _build_uk_weight_replacement_region and _require_certified_uk_weight_matrix, which #881 deleted.

How it was resolved.

• main's code is kept, including its UK area routing helpers.
• The UK region type and input dataset entity data fixes above went with the deleted functions.
• The compiled reform None check and the compiled policy variable rename (runtime_policy) still apply.
• The second commit is now "Type the simulation runtime's compiled-policy reads".
• The baseline and the registry count were regenerated from a Pyright run, not merged by hand, and docs/generated/rules.md was regenerated with python -m scripts.rules render.

Commit	Errors fixed	Baseline after
Open the simulation runtime's spans with the typed span scope	0	68
Type the simulation runtime's compiled-policy reads	2	66

Baseline at this PR's head: 66 entries (main has 68).

Testing

• python -m scripts.rules pyright --sync: the executor passes with its 66 baselined errors; every other package matches its rule (gateway 0, apis-integ 4 baselined, the libraries and entry 0).
• Executor unit tests: 1865 passed, 22 skipped.
• Quality guards, the rules tool's tests (191) and ruff 0.15.22 format and check: pass. No new noqa, type: ignore or pyright: ignore.

This PR is now the bottom of the stack, so its CI runs in full.

🤖 Generated with Claude Code