The Master Key Hypothesis: Unlocking Cross-Model Capability Transfer via Linear Subspace Alignment