Zero-Direction Probing: A Linear-Algebraic Framework for Deep Analysis of Large-Language-Model Drift