feat: multi-component hydrograph loss function

[taddyb]

Summary

• Replaces the single MSE loss with a configurable multi-component hydrograph loss (hydrograph_loss) targeting peak amplitude, baseflow, timing, and overall fit — each per-gage normalized to prevent large basins from dominating
• Adds LossConfig (Pydantic) to ExperimentConfig with sensible defaults so existing configs work unchanged
• Adds per-component loss logging in train.py for diagnostics during training
• Adds comprehensive unit tests for the loss function and an integration test for the training loop

Changes

• src/ddr/validation/losses.py (new): hydrograph_loss with four components — overall MSE, peak regime (P98), baseflow regime (P30), and temporal gradient loss. Per-gage loss capped at 10 to prevent variance-collapse blowup.
• src/ddr/validation/configs.py: LossConfig with weights, percentiles, and eps. Added as loss field on ExperimentConfig with defaults.
• src/ddr/validation/__init__.py: Re-exports hydrograph_loss.
• scripts/train.py: Uses hydrograph_loss instead of mse_loss; logs individual component values.
• benchmarks/: Minor config/import alignment.
• tests/validation/test_losses.py (new): Unit tests for all loss components.
• tests/integration/ (new): End-to-end training loop integration test.

Test plan

• uv run pytest tests/validation/test_losses.py — unit tests for loss components
• uv run pytest tests/integration/test_training.py — integration test for training loop
• Verify existing configs still work (LossConfig defaults applied automatically)
• Training run on HPC to validate gradient flow through all components

🤖 Generated with Claude Code