Consensus-Target Linear Shrinkage: The Geometry of In-Context Learning in the Gradient-Gram Space

In-context learning (ICL) performs consensus-target linear shrinkage on the 
task-gradient Gram matrix: G^P = α_P·J + (1-α_P)·G^r. This implies an 
Orthogonal Contraction Law and an Effective-Rank Entropy Bound. A 65-point 
sweep confirms α_P as the master variable (per-task R² > 0.98). Cross-model 
validation on 8 architectures (117M-2.7B) shows prediction error below 3%. 
Combined with the ECC decomposition, this yields the Adaptation Cone Law 
constraining intelligence-improving adaptations.