Revert to v0.5.1 baseline with Leopold & Maddock physics

[taddyb]

Summary

• Revert to v0.5.1 baseline: removes phi_kan bias correction, SWOT geometry, temporal KAN, forcing readers, and staged training
• Leopold & Maddock power law: derives top_width and side_slope per-timestep from p_spatial * depth^q_spatial instead of learning them independently. Adds _apply_data_override() for Lynker/partial-coverage blending
• p_spatial replaces top_width/side_slope as a learnable KAN output (width coefficient, log-space [1, 200])
• In-place autograd fix: scatter_add → clamp no longer double-writes to output tensor
• MSE → MAE as default training loss
• CUDA memory cleanup: explicit del + empty_cache() after each mini-batch

Test plan

• All 429 tests pass (pytest tests/ --ignore=tests/references)
• Run scripts/train_and_test.py with merit_training_config.yaml (MERIT path: power-law derived geometry)
• Run scripts/train_and_test.py with lynker_train_and_test_config.yaml (Lynker path: data override)
• Verify gradient flow through p_spatial → routing → loss

🤖 Generated with Claude Code

[taddyb]

results from MERIT train/test

[2026-03-14 06:27:34,987][__main__][INFO] - Evaluating with checkpoint: _ddr-v0.5.2.dev2+g21a3a96b5-merit-training_epoch_5_mb_35.pt
[2026-03-14 06:27:35,047][ddr.io.readers][INFO] - Reading icechunk store from local disk: /home/tbindas/projects/ddr/data/merit_dhbv2_UH_retrospective
[2026-03-14 06:27:35,259][ddr.io.statistics][INFO] - Reading Attribute Statistics from file: merit_attribute_statistics_merit_global_attributes_v2.nc.json
[2026-03-14 06:27:38,155][ddr.io.readers][INFO] - Reading icechunk store from local disk: /mnt/ssd1/data/icechunk/usgs_daily_observations
[2026-03-14 06:27:38,190][ddr.geodatazoo.merit][INFO] - Filtered 352 gages with DA_VALID=False
[2026-03-14 06:27:40,106][ddr.geodatazoo.merit][INFO] - Filtered 494 headwater gages with no upstream connectivity
[2026-03-14 06:27:40,106][ddr.geodatazoo.merit][INFO] - Gages mode: 2365 gauged locations
[2026-03-14 06:27:44,530][ddr.geodatazoo.merit][INFO] - Created gages adjacency matrix of shape: torch.Size([64892, 64892])
[2026-03-14 06:27:44,533][ddr.scripts_utils][INFO] - Loading spatial_nn from checkpoint: _ddr-v0.5.2.dev2+g21a3a96b5-merit-training_epoch_5_mb_35
[2026-03-14 06:34:19,019][ddr.validation.utils][INFO] - ----------------------------------------
[2026-03-14 06:34:19,019][ddr.validation.utils][INFO] - Metric     |         Mean |       Median
[2026-03-14 06:34:19,019][ddr.validation.utils][INFO] - ----------------------------------------
[2026-03-14 06:34:19,019][ddr.validation.utils][INFO] - NSE        |       0.5738 |       0.7332
[2026-03-14 06:34:19,019][ddr.validation.utils][INFO] - RMSE       |      17.3540 |       8.0274
[2026-03-14 06:34:19,019][ddr.validation.utils][INFO] - KGE        |       0.6425 |       0.7407

[taddyb]

Lynker hydrofabric performance

[2026-03-14 13:11:44,633][ddr.validation.utils][INFO] - ----------------------------------------
[2026-03-14 13:11:44,633][ddr.validation.utils][INFO] - Metric     |         Mean |       Median
[2026-03-14 13:11:44,633][ddr.validation.utils][INFO] - ----------------------------------------
[2026-03-14 13:11:44,633][ddr.validation.utils][INFO] - NSE        |       0.5213 |       0.7037
[2026-03-14 13:11:44,633][ddr.validation.utils][INFO] - RMSE       |      16.3834 |       7.0815
[2026-03-14 13:11:44,633][ddr.validation.utils][INFO] - KGE        |       0.6354 |       0.7487
[2026-03-14 13:11:44,633][ddr.validation.utils][INFO] - ----------------------------------------
[2026-03-14 13:11:44,633][__main__][INFO] - Test run complete. Please run examples/eval/evaluate.ipynb to generate performance plots / metrics
[2026-03-14 13:11:44,662][__main__][INFO] - Cleaning up...
[2026-03-14 13:11:44,663][__main__][INFO] - Total time: 84.92 minutes