Revamp Section 6 presence-background evaluation

[mrecos]

Summary

• make common 7 × 7 focal support the primary Section 6 design for sites, training backgrounds, and mapped availability
• score one fixed uniform sample of valid raster anchors in every spatially grouped fold
• transform held-out predictions to fold-specific availability percentiles and pool every bag exactly once per repeat
• lead with mapped-area capture, lift, gain, Continuous Boyce, and held-out site ranks; retain ROC AUC and PR AUC as secondary diagnostics
• rebuild notebook 05 as a streamlined report driven by one tracked runner/config/result contract
• add fold-safe availability maps, geometry-only controls, M0–M3 and LR/RF comparisons, compute diagnostics, and optional 9 × 9, 11 × 11, and original-irregular support sensitivities
• speed nested focal extraction by reading the largest support once and reuse raster handles for batched availability windows

Scientific rationale

Section 6 is presence–background, not presence–confirmed-absence. The constructed background prevalence and a binary threshold are therefore not the main prospection questions. This revision evaluates how efficiently a fixed mapped-area budget captures held-out sites and whether sites concentrate toward higher mapped suitability.

Each fold gets the same availability anchors. Scores from independently fitted folds are placed on a common availability-rank scale before pooling. Tied thresholds explicitly record the achieved mapped fraction and use that achieved fraction—rather than the nominal budget—as the denominator for lift and gain, which avoids favoring discrete models such as random forests.

The geometry-only negative control remains visible. Strong geometry performance under common focal support is treated as a design warning. No method is promoted or removed from this single physio-shed.

Notebook workflow

notebooks/05_section6_model_validation.ipynb now follows one reading order:

1. configuration and reproducibility contract
2. data/support audit and leakage checks
3. geometry gate
4. mapped-area capture/lift/gain
5. Boyce and held-out site ranks
6. secondary ROC AUC/PR AUC
7. fold-safe maps with training and held-out sites
8. compute diagnostics
9. optional support sensitivity
10. structured, non-promotional interpretation

The notebook runs the primary evaluation automatically when no result exists and loads cached ignored output on later runs. benchmarks/build_section6_evaluation_notebook.py regenerates it.

Validation

• 148 pytest tests passed
• Black passed on all CI-tracked Python paths
• Ruff passed
• Mypy passed
• pyproject.toml parsed successfully
• real riverine preparation passed: 142 focal sites, 142 matched focal backgrounds, 1 explicit site exclusion, and 1,000 availability anchors
• bounded real-data riverine LR evaluation passed for 10 folds / 2 pooled repeats
• result schema validation passed
• all 25 notebook cells parsed and the full notebook report executed successfully against the bounded real-data result

Dependency

This is a stacked draft PR targeting codex/representation-extensions because PR #9 has passed CI but still awaits explicit merge approval. After #9 is merged, retarget this PR to main; its diff will then contain only the evaluation revamp.

Untracked * 2.* sync-copy artifacts were deliberately excluded.