Loss of cached computation when prompt structure changes, requiring recomputation of affected tokens.