Fix sheaf-theoretic rigor: Laplacian restriction maps, terminology al…

[MikeHLee]

…ignment, and neuro validation

• Add matrix property to Transformation class for explicit restriction map representation
• Fix compute_sheaf_laplacian to use actual restriction maps R instead of identity matrices
  • Off-diagonal blocks: L_ij = -R_ij (not -I)
  • Diagonal blocks: L_ii = Σ_k R_ik^T @ R_ik
  • Enables detection of twist inconsistencies (rotations, calibrations)
• Correct docstring terminology in transforms.py
  • text_to_tokens, image_to_patches: Restriction maps (Global → Local)
  • tokens_to_text, patches_to_image: Gluing/Extension maps (Local → Global)
• Add validate_signal_assumptions() to BrainSheaf
  • Checks variance disparity (>10x warns about z-scoring needed)
  • Checks mean centering across regions
  • Auto-validates on first detect_dissonance call
  • Prevents false dissonance from high-variance regions

Addresses conceptual gaps identified in math rigor review.