Use modal deploy + spawn to decouple pipeline from GitHub runner

[baogorek]

Summary

modal run --detach streams logs to the GitHub Actions runner and dies when the runner timeout fires, killing the Modal pipeline mid-run. The pipeline takes 10+ hours, well past GitHub's 6-hour max.

Fix: Deploy the app to Modal (making it persistent), then .spawn() the run_pipeline function as a fire-and-forget call. The GitHub Actions step finishes in under a minute while the pipeline runs on Modal untethered for as long as it needs (up to Modal's 24-hour function timeout).

Fixes #636

Before

modal run --detach modal_app/pipeline.py::main $ARGS
# CLI streams logs → runner timeout kills pipeline

After

modal deploy modal_app/pipeline.py
# App is now persistent on Modal

python -c "run_pipeline.spawn(branch='main', ...)"
# Fire-and-forget → step exits immediately

Test plan

• Merge and verify the GitHub Actions step completes in ~1 minute
• Verify the pipeline appears and runs on the Modal dashboard

🤖 Generated with Claude Code

[juaristi22]

Tested the deploy + spawn approach locally and it works correctly:

1. modal deploy modal_app/pipeline.py — app registered successfully as policyengine-us-data-pipeline
2. run_pipeline.spawn(...) — returned immediately with a function call ID, pipeline running on Modal independently

Minor note: The deploy step took ~22 minutes locally (first-time image build). The workflow timeout-minutes is currently set to 10, which would kill the deploy before it reaches the spawn call — the same failure mode as the original issue. Should bump to 30 to handle worst-case deploy times (subsequent deploys will be faster due to Modal image caching, but first deploys or cache misses need headroom).

[juaristi22]

Fix looks correct — deploy + spawn approach works as expected. Tested locally end-to-end. Just needs the timeout bump to 30 minutes.