SEM CFA results in error for longitudinal data

#1
I am trying to build a longitudinal measurement model of a "strategy
construct" using SEM.
As long as I am building the model for each year individually, I have no
problem and Stata's SEM converges. I use the following code:

_________________________________________________________________________
sem (diff2011 -> wmdiff_12011) (diff2011 -> wpdiff_12011) (diff2011 ->
wpdiff_22011), iterate(1000) latent(diff2011 ) cov( diff2011@1)
nocapslatent
_________________________________________________________________________

As you can see, I set the variance of the latent construct to 1 in order
to avoid using a Marker variable.

My problem arises when I want to create a longitudinal model. As the
indicators are allowed to covary over time, same as the latent construct,
I specifically model this covariance, using the following code:

_________________________________________________________________________
sem (diff2010 -> wmdiff_12010) (diff2010 -> wpdiff_12010) (diff2010 ->
wpdiff_22010) (diff2011 -> wmdiff_12011) (diff2011 -> wpdiff_12011)
(diff2011 -> wpdiff_22011), covstruct(_lexogenous, diagonal) iterate(1000)
latent(diff2010 diff2011 ) cov( e.wmdiff_12010*e.wmdiff_12011
e.wpdiff_12010*e.wpdiff_12011 e.wpdiff_22010*e.wpdiff_22011 diff2010@1
diff2010*diff2011 diff2011@1) nocapslatent
_________________________________________________________________________

Again I set the variances of the latent construct equal to 1 (but even if
I don't I get the error).

The estimation quickly stops with the error message "Estimation failed"
and in the results window I get the message that "initial values not
feasible
r(1400);
"

Does anyone know why this happens or how to do it properly?