To the best of my knowledge, as items (or observed variables) are actually measured by the researcher ( not the Latent Variables), thus for identifying the characteristics of a normal distribution of data it is better to do a normality test separately for each items. For instance, the normality test for a computed variable (A) indicates normal distribution while we dont know whether all items of the varable (A) are normally distributed or not. Therefore the normally distributed items shows the high degree of accuracy of data.