[vgvr0]

Summary

Fix SMAPE handling of zero and missing forecasts across pandas and Polars.

Previously, fill_null(0) conflated two different cases:

• y = 0 and y_hat = 0, which should contribute zero error.
• Missing forecasts, which should remain missing and be excluded from the mean.

This also produced different behavior between pandas and Polars.

Changes

• Handle a zero SMAPE denominator explicitly as 0.0.
• Preserve missing forecasts instead of turning them into perfect predictions.
• Align pandas and Polars behavior.
• Add regression tests covering:
  • 0/0
  • partially missing forecasts
  • completely missing forecasts
  • ordinary forecasts

Validation

• tests/test_losses.py: 122 passed, 8 skipped
• Full test suite: 1746 passed, 9 skipped
• Coverage: 88.94%

Fixes #285

[Mariana Menchero García (MMenchero)]

Thank you for your contribution vgvr0.