test_budget_window_semi_integration is a unit test, not an integration test

[MaxGhenis]

Summary

tests/test_budget_window_semi_integration.py is labeled "semi-integration" but exercises nothing real. It uses MockModal, MockDict, a mocked time.sleep, and a fake FastAPI TestClient. The file never touches Modal, network, or any external dependency; it is strictly a unit test and should be labeled as such. There is also no genuine integration coverage for the budget-window orchestration path.

Location

projects/policyengine-api-simulation/tests/test_budget_window_semi_integration.py:1-60

What goes wrong

"""Semi-integration coverage for budget-window gateway and worker seams."""

from fastapi.testclient import TestClient

import src.modal.budget_window_batch as batch_module
import src.modal.budget_window_scheduler as scheduler_module
...

class MockDict:
    def __init__(self, data: dict):
        ...

The test substitutes every external seam (Modal Dicts, FunctionCall, time.sleep) and operates entirely in-process. That is the definition of a unit test, not semi-integration. Calling it "semi-integration" gives false confidence that the budget-window path has real-world coverage — in reality the scheduler's interaction with Modal's actual RPC and Dict semantics (stale reads, timeouts, FunctionCall.from_id edge cases) is not exercised at all.

Suggested fix

• Rename to test_budget_window_scheduler_unit.py (or similar) to reflect reality.
• Add a genuine integration test that runs the gateway + a Modal ephemeral app (e.g. via modal run with a lightweight run_simulation stub) and asserts end-to-end status transitions.
• Gate the integration test on a @pytest.mark.integration marker and a Modal credential, skipped in local/no-credential CI runs.

Severity

Low. Test-quality / coverage-signal issue, not a behavior bug, but meaningful because the budget-window path is new and unauthenticated.