Clean up baseline persistence and comparison tests

[charlesmartin14]

Scope

This is the narrowly scoped minor-cleanup pass from the baseline code review. It deliberately does not change optimizer mathematics, hyperparameters, experiment selection policy, notebook code, committed output notebooks, or existing baseline results.

Changes

• add atomic CSV and compressed-NPZ replacement helpers for MNIST progress artifacts
• use those helpers for per-epoch performance, spectral, WeightWatcher-detail, optimizer-group, and ESD-history persistence
• remove the ineffective locals() assignment loop from MNIST resume handling while retaining the explicit truncation logic
• tighten validation for seeds, optional evaluation limits, SGD dampening, Muon parameter names/Nesterov momentum, and WeightWatcher bounds/method names
• rename the comparison result field to paired_terminal_differences, with a backward-compatible paired_final_differences property
• add synthetic MNIST comparison tests, including complete-suite execution and duplicate-grid rejection
• add atomic-persistence and configuration-validation tests

Validation

• git diff --check
• python -m compileall -q baseline/rg_baselines baseline/tests
• python -m unittest discover -s baseline/tests -v
• 62 tests passed
• dependency consistency passed with pip check

The temporary workflow used to apply and validate the patch was removed; the PR's net diff contains only the six intended baseline source/test files.

Deferred to the next review

The larger findings remain intentionally untouched: one-head completed-run fingerprint validation, stricter test-holdout policy, convergence censoring, paired one-head contrasts, interruption-equivalence testing, Muon known-answer tests, and the ViT monkeypatch refactor.