Add early stopping and regularization study

[almondsun]

Summary

• add validation-based early stopping with inspectable stop state and preserved best/final checkpoint semantics
• fail fast on non-finite losses before optimizer mutation and require standards-compliant JSON artifacts
• compare ByteBPE512 early stopping with and without AdamW weight decay 0.01
• record milestone 022 results and extend training theory/documentation

Results

• early stopping reproduces best BPC 2.008269 at step 1750 and stops at step 2500
• runtime falls from 422.7s to 209.3s and stopped-final BPC improves from 2.244955 to 2.055447
• weight decay 0.01 reaches best BPC 2.008018; the 0.000251 difference is effectively neutral

Validation

• uv run --frozen --extra dev make check (121 tests, 91.18% coverage)
• uv run --frozen --extra dev make audit
• two controlled training runs plus best/final greedy and seeded generation diagnostics
• post-change correctness and security audits; all findings resolved

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.