Geometry predictor script with global KAN + slope floor fix

[taddyb]

Summary

• New scripts/geometry_predictor.py — Hydra script that predicts KAN parameters (n, p, q) for all 2.9M global MERIT reaches and computes trapezoidal geometry statistics (min/max/median/mean for depth, top_width, bottom_width, side_slope, hydraulic_radius, discharge) for 346K CONUS reaches with streamflow data. Non-CONUS reaches get KAN params with NaN geometry stats.
• New src/ddr/geometry/statistics.py — Pure compute_geometry_statistics() function (no I/O) that takes KAN params + daily accumulated discharge and returns per-reach temporal statistics.
• Slope minimum raised from 0.0001 to 0.001 — The old floor caused 200m+ depth predictions on flat high-Q reaches (Mississippi delta). Updated across configs.py, predictor.py, and all test fixtures. Model will need retraining with the new floor.
• Simplified example notebooks — Both geometry_predictor.ipynb and plot_parameter_map.ipynb are now thin consumers that load the script's .nc output and plot. All inline routing code removed.
• New config/merit_geometry_config.yaml — Hydra config for WY2000 (1999/10/01–2000/09/30) with 30-day batches.

Files changed

File	Change
scripts/geometry_predictor.py	New Hydra script
src/ddr/geometry/statistics.py	New pure computation module
src/ddr/geometry/__init__.py	Export compute_geometry_statistics
src/ddr/validation/configs.py	Slope minimum 0.0001 → 0.001
src/ddr/geometry/predictor.py	Slope fallback 0.0001 → 0.001
config/merit_geometry_config.yaml	New Hydra config
examples/merit/geometry_predictor.ipynb	Simplified to load .nc + plot
examples/merit/plot_parameter_map.ipynb	Simplified to load .nc + plot
tests/geometry/test_geometry_stats.py	4 unit tests for statistics
tests/geometry/test_predictor.py	Updated slope fixture
tests/routing/test_utils.py	Updated slope fixture
tests/validation/test_configs.py	Updated slope default test
tests/benchmarks/conftest.py	Updated slope fixture

Test plan

• uv run pytest tests/geometry/test_geometry_stats.py — 4 pass
• uv run pytest tests/geometry/ tests/validation/test_configs.py tests/routing/test_utils.py — 53 pass
• Full script run produces 353 MB NetCDF with 2.9M reaches (2.5 min on CPU)
• Retrain model with new slope floor (0.001) before production use

🤖 Generated with Claude Code