Add tiling artifact diagnostics eval

[Hgherzog]

Summary

• Add TILING_DIAGNOSTICS eval mode that detects spatial tiling/striping artifacts via row/col variance anisotropy and 2D FFT axis energy, plus PCA RGB visualization logged to W&B
• Add TaskType.DIAGNOSTIC to cleanly separate diagnostic datasets from real segmentation tasks, preserving spatial dims without requiring probe_lr/ft_lr
• Add per-resolution pretrain subset configs (pretrain_subset_64, pretrain_subset_128) with backward-compat pretrain_subset alias
• Wire tiling diagnostics into in-loop evals (script.py), standalone evals (all_evals.py), and checkpoint sweeps (checkpoint_sweep_evals.py via TILING_DIAGNOSTICS_ONLY)
• Add evaluation docs for both embedding diagnostics and tiling diagnostics (metrics, launch commands, interpretation)

Test plan

• Unit tests for compute_tiling_artifact_metrics (stripe detection, FFT period, isotropic baseline, edge cases)
• Unit tests for pca_rgb_image (shape, dtype, value range)
• Existing embedding diagnostics + config tests still pass
• Smoke test in-loop eval on a real checkpoint to verify W&B logging

Made with Cursor

---

Note

Medium Risk
Adds a new evaluation mode and dataset/task-type handling in the core eval pipeline, plus new FFT/PCA computations and W&B image logging, which could affect eval stability/perf and downstream task selection logic.

Overview
Adds a new tiling artifact diagnostics evaluation path (EvalMode.TILING_DIAGNOSTICS) that computes striping/tiling signals from spatial embeddings (row/col variance ratio, FFT axis energy, dominant period) and optionally logs a PCA-RGB visualization to W&B.

Introduces TaskType.DIAGNOSTIC and per-resolution pretrain_subset_64/pretrain_subset_128 configs (with pretrain_subset as an alias) so diagnostic datasets preserve spatial dimensions and can be run without probe/finetune learning-rate requirements; updates the pretrain-subset loader to derive hw_p from config and to return per-pixel dummy labels.

Wires these diagnostics into in-loop training evals (scripts/official/script.py), standalone task registries (internal/all_evals.py), and checkpoint sweep evaluation selection/logging (checkpoint_sweep_evals.py via TILING_DIAGNOSTICS_ONLY), and expands docs/tests accordingly.

^{Reviewed by Cursor Bugbot for commit df61340. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes and found 2 potential issues.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit df61340. Configure here.}

[cursor]

You have used all of your free Bugbot PR reviews.

To receive reviews on all of your PRs, visit the Cursor dashboard to activate Pro and start your 14-day free trial.

[cursor]

You have used all of your free Bugbot PR reviews.

To receive reviews on all of your PRs, visit the Cursor dashboard to activate Pro and start your 14-day free trial.