Fix coordinate_publish import path that pulls in policyengine_core on the local-area image

[MaxGhenis]

Problem

On April 3, 2026, the post-merge pipeline discussed on PR #684 failed in coordinate_publish with:

ModuleNotFoundError: No module named 'policyengine_core'

The immediate import site is in modal_app/local_area.py:

from policyengine_us_data.calibration.calibration_utils import STATE_CODES

That import is too heavy for the local-area Modal image. policyengine_us_data/__init__.py eagerly imports .datasets, which pulls in modules that depend on policyengine_core and policyengine_us. The coordinate_publish image does not include those packages, so the run fails before H5 generation can proceed.

Root cause

coordinate_publish only needs STATE_CODES, but it imports that constant from calibration_utils, which transitively imports the full package dependency chain.

Smallest safe fix

Move STATE_CODES to a lightweight module with no heavy package imports, or inline/use an already-lightweight source for that mapping, and update modal_app/local_area.py to import from there instead.

The goal is to keep the coordinate_publish path importable on the local-area image without requiring policyengine_core or policyengine_us.

Acceptance criteria

• coordinate_publish no longer imports policyengine_core transitively when building work items
• The local-area Modal image can execute the work-item discovery path without ModuleNotFoundError
• The fix does not broaden the local-area image dependency set

Context

• Discussed in PR Scale H5 pipeline to 50 workers at 1 CPU each #684 on April 3, 2026
• Separate from the worker-count default change in that PR
• Potentially related but distinct from version-skew concerns observed in the same post-merge run