What does the SEM command (for a simple latent measurement model) do with missing values when 'predict' into a linear variable?

[Eric ES]

If I use the SEM command to construct a fairly simple latent measurement model and then use 'predict' to construct a linear variable out of the new latent variable, it creates real values for the missing observations. But the SEM command is supposed to use listwise deletion for missings (or so I thought) when not specifying FIML. What am I missing here? I have attached a log file with just these actions (as PDF). See the number of obs for the model and then the final number at the bottom of the 'tab1 *, m' at the end of the log. That final number is my total sample. Stata SEM Help Log file (smcl).pdf

[Eric ES]

Note, I realize I actually did use FIML in this example, but it clarifies that the 72 all-missing obs were excluded. But they clearly have data for the predicted variable.