Add Level 0 three-by-three epoch diagnostics notebook - #123

What changed

Adds level_0_wwpgd/notebooks/04_repeat_3x3_epoch_diagnostics.ipynb to rerun the original isolated Level 0 experiment as a matched three-seed comparison:

• AdamW baseline: seeds 1337, 2027, 4099
• AdamW + stock WWPGD: the same three seeds
• frozen 4-block, 4-head, width-128, context-256 architecture
• existing 2,000-step training protocol and stock WWPGD event projection

Notebook outputs

The notebook runs the repository's existing paired runner, verifies the frozen configuration before launch, and plots all curves against token-budget-equivalent training epoch (tokens_seen / train_split_tokens). It covers:

• train, validation, and test cross-entropy
• train, validation, and test perplexity
• train/validation token accuracy
• post-hoc test top-1 and top-5 token accuracy
• test bits per token
• validation/test loss and accuracy gaps
• layerwise WeightWatcher alpha
• alpha-fit KS D, sigma, fitted-tail support, detX retained count, and ERG gap
• requested, applied, maximum, and cumulative WWPGD relative Frobenius correction
• paired final-checkpoint summaries by seed

Test curves are evaluated after training from chronological saved checkpoints using a fixed per-seed test probe. They do not affect optimization or checkpoint selection. Analysis CSVs and PNGs are written under the selected paired result root.

Validation

• notebook passes nbformat schema validation
• every Python code cell passes ast.parse
• branch differs from main by this notebook only

The six full training runs were not executed in this environment because the prepared FineWeb-Edu token data and target accelerator are not present.