My recent  concern in the data analysis process is whether to include  factor covariances and errors of measurement in the SEM analysis. I have run CFA with these conditions and the model show appropriate goodness of fit indices, but since SEM  only tells about the relationships between latent variables I am doubtful if we can put those conditions here as well. Moreover, we also rely on modification indices for model improvement. Again there are two questions: 

1) If I am continuing the SEM analysis in CFA model i.e the model showing factor covariances and errors of measurement the goodness of fit indices (GFI, CFI, IFI and TLI) is indicating a good fit. But do I have to give theoretical arguments of covariances between the factors? I mean there are approximately 40-42 such covariances and many of them do not fit with theory.

2) And,  If I am conducting the SEM analysis without taking factor covariances and errors of measurement, the goodness of fit indices (GFI, CFI, IFI and TLI) is indicating a poor fit. However, if I am making changes as per Modification Indices (that is theoretically fitting 14-16 factor covariances), the fit indices values are just above the threshold values and not meeting RMR and SRMR conditions.

Please tell which approach is correct