Your first assumption is correct, that the covariates vary across subjects but not condition. (not that I explained it terribly well)
The second level scans represent intra-subject contrasts (emotion vs. shape) so I think you are right that the subject effects should have been subtracted out.
Thanks for helping to clarify the models - sounds like there is some utility in both which I'll have to give some more thought.
For model 2, am I correct in assuming that if I select either a single condition, or a single covariate interaction term in the contrast manager, that I will be looking at the condition x covariate interaction?