feat: add BoundaryNudgingLayer for regional-global blending

[AswaniSahoo]

Pull Request

Description

This PR adds BoundaryNudgingLayer, which blends regional predictions with global context at domain boundaries using a relaxation zone.

It builds on RegionalForecaster from #221 to support boundary conditioning for issue #3. The layer is opt-in and does not change behavior when disabled.

The layer includes:

• Distance-based relaxation prior using haversine distance from region centroid
• Learned MLP correction on top of the geometric prior
• Smooth blending: alpha=0 at center (trust regional), alpha=1 at edges (trust global)
• Integration into RegionalForecaster.forward() after the residual connection
• Two new config fields: enable_nudging (default False), nudging_hidden_dim

Related to #3

How Has This Been Tested?

The following tests were added and run using pytest:

• tests/test_regional_forecast.py
  • Nudging disabled does not change output (backward compat)
  • Nudging enabled with no global context produces valid output
  • Global context changes output when nudging is enabled
  • Relaxation weights are in [0, 1] with max at farthest point
  • Gradients flow through the nudging layer

Commands run:

• .\.venv\Scripts\python.exe -m pytest tests/test_regional_forecast.py -v

  • 13 passed, 1 warning

• .\.venv\Scripts\python.exe -m ruff check graph_weather/models/regional_forecast.py tests/test_regional_forecast.py

  • Passed

If your changes affect data processing, have you plotted any changes? i.e. have you done a quick sanity check?

• Not applicable - no data processing changes

Checklist:

• My code follows OCF's coding style guidelines
• I have performed a self-review of my own code
• I have made corresponding changes to the documentation
• I have added tests that prove my fix is effective or that my feature works
• I have checked my code and corrected any misspellings

[AswaniSahoo]

@jacobbieker This implements the boundary nudging approach we discussed on Issue #3 .