Say I have 10 strategy parameters and I run an optimization involving only 5 of them. Then when I set the Preferred Values, it seems that the parameters that were not involved in the optimization are deleted from the code, and I need to retype them in. What am I doing wrong?

It might help if you describe what you're doing step by step.

Open strategy (code).
Select Dataset.
Click Optimize.
In Optimization Control tab, select one of the parameters, then click Remove Selected Parameter. Repeat for all params to not be used in the optimization.
Select Genetic optimization.
Begin optimization.
Run to completion.
Results > Double- click Net Profit to sort for max Net Profit.
Select top row (max Net Profit), right click, Set these Parameter Values as Default for the Strategy.
POOF! Wealth Lab deletes the parameters that I removed from the optimization from my code, and I have to type their declarations and values back in manually.

QUOTE:
Wealth Lab deletes the parameters that I removed from the optimization from my code

Yes, it is expected. Until Fidelity agrees to introduce a checkbox system to enable/disable parameters on-the-fly, removing selected parameters removes them from the code.

You might want to call your Fidelity rep and tell that you like this to be fixed (QC 98351):

(98351) The Optimization control lacks the ability to selectively omit Strategy Parameters from optimizations

In fact, there's nearly a dozen bugs and enhancement requests for the Optimizer registered in the system. If everything goes as planned, chances are that some of them they may get resolved in 6.6 (I'll keep my fingers crossed).

QUOTE:
Until Fidelity agrees to introduce a checkbox system to enable/disable parameters on-the-fly, removing selected parameters removes them from the code.
You might want to call your Fidelity rep and tell that you like this to be fixed (QC 98351)

I too have encountered this problem in WLP 6.8.10. In practice, out of six parameters in my most powerful strategy, four of those parameters remain constant for 95% of the symbols. So I only want the optimizer to optimize two of the six parameters, and only for about 10% of the dataset symbols.

In practice, I quit using the optimizer altogether. It doesn't work all that well (I can optimize the really exceptionally tough cases by hand better), and it has too many limitations. In have thought about placing the exceptionally tough cases (5-10%) in a special dataset for manual Preferred Value management, but that would cause me to double the number of datasets I have to manage.

Also, when managing the parameters of the tough cases manually, I've discovered I occasionally revert back to the default values over time. What's really needed is a Wiener optimizer with a feedback loop to dynamically adjust the more unstable parameters in the tough cases. But we are dealing with nonlinear behavior here, so I doubt a Wiener optimizer would ever be able to converge on a useful solution. It's a tough numerical problem.