Phase 2 Person 4: full-scale ablation fold-in + DeepLab multi-seed

[LasanaPahanga]

Summary

• Add Phase2/Lasana-Person4/: fold Chanupa U-Net + Kalana SegFormer-B0 + Dinura l2_1_mse attention winner into one full-scale table; DeepLabV3+ multi-seed (42/43/44) mean±std pipeline.
• Review fixes: project-relative checkpoint paths (no absolute C:\Users\...); single DeepLab seed-42 Dice (0.7821 from deeplab_multiseed.json); local aggregate_mean_std with sample std (ddof=1) and bare values for n=1 (no ± 0.0000).
• Boundary Loss ablation row stays pending until Person 5 (Dhinanjaya) wires Boundary Refinement and hands off a baseline_comparison.csv row — branch is mergeable without it. TODO in fold_full_scale_results.py names the expected upstream path.

Multi-seed convention

Persons 1/2/3 GPU multi-seed reporting must use sample std (ddof=1) to match this branch.

Test plan

• cd Phase2/Lasana-Person4 && python -m unittest tests.test_fold_results -v (6/6)
• python fold_full_scale_results.py
• python train_deeplab_multiseed.py --skip-train
• no C:\Users under Phase2/Lasana-Person4/; no ± 0.0000 in results/
• Reviewer: confirm DeepLab row caption excludes full held-out set; confirm attention row is l2_1_mse