Add lightweight metrics and WeightWatcher audit

[charlesmartin14]

Summary

• fix the single-run MuonClip Jacobian audit so completed-run provenance is validated from the persistent run while model references come from the verified tail cache
• normalize Papermill JSON-list and boolean strings in the one- and three-seed weight-quotient notebooks
• add notebook 21, a lightweight CSV-only dashboard for test accuracy/loss and per-layer WeightWatcher alpha (raw versus fix_fingers=clip_xmax)
• persist sanity checks, fit availability, source rows, plots, an audit summary, and strict provenance
• document the exact command-line invocation

Performance

Notebook 21 loads no checkpoints, runs no Jacobians or child notebooks, and performs no new WeightWatcher fits. A synthetic completed-run execution produced all expected outputs in about three seconds.

Tests

• python -m unittest baseline.tests.test_source_syntax baseline.tests.tangent_rg.test_notebook_contract baseline.tests.tangent_rg.test_weight_quotients baseline.tests.tangent_rg.test_weight_quotient_resume -v (37 tests)
• git diff --check
• synthetic end-to-end execution of every code cell in notebook 21 with all expected output artifacts and zero failed sanity checks