Add Muon weight quotient checkpoint notebooks

[charlesmartin14]

Summary

• add post-facto one-seed and three-seed Muon/MuonClip weight-quotient notebooks over the verified final-100 checkpoint cache
• materialize five nonlinear spectral inverse models as full-shape canonical representatives of the diagnostic O(out) x O(in) quotient:
  1. scalar Gram ridge subtraction
  2. blockwise singular counterterms
  3. anchor-frozen Feshbach downfolding
  4. rectangular empirical D-transform deconvolution
  5. discarded-bulk-calibrated monotone MP shrinkage
• retain raw weights, midpoint ECS truncation, and uniform singular translation as explicit controls
• run every transformed matrix through the existing dual WeightWatcher path, both raw and fix_fingers="clip_xmax"
• add exact result-grid, ESD identity, checkpoint provenance, cross-seed provenance, resumability, and completion-manifest audits
• add Student-t 95% intervals for exactly three complete seeds; the one-seed notebook deliberately has no error bars
• document that this is a spectral diagnostic quotient, not a ReLU network symmetry or an exact quotient by arbitrary unknown Muon histories

Verification

• 35 focused source-syntax, generated-notebook, algebra, WeightWatcher-row-contract, and resumability tests pass
• mocked matrix replacement -> actual ESD -> dual WeightWatcher -> resumable reload integration passes
• generated notebooks reproduce exactly from build_notebooks.py
• git diff --check passes

The local environment does not contain the real final-100 caches or the full WeightWatcher/PyTorch runtime, so full checkpoint execution remains the baseline-run follow-up; the dependency-light integration and all static/algebraic contracts pass.