I am just a Ph.D. student. My main research area is not Bayesian statistics. My suggestion would be
(1) Model selection is a big research area. What type of distribution assumption should be adopted or what predictors need to be used, all based on your data type. You can use DIC to pick an "optimal" distribution assumption. Also, based on MCMC sample that has been generated by Openbugs, calculate 95% HPD interval for each coefficient(beta_i) of covariate. Get rid of insignificant covariates (backward or forward). Refit the model.
(2) Posterior predictive distribution can be obtained by integration. You probably want to do that by hand at first. If you can find that, it's perfect. If not, through sampling technique(monte carlo idea), you still can get numerical approximation of p(Y_new|X_new, Y_observe, X_design matrix ). Then find
E(Y_new|X_new, Y_observe, X_design matrix) either theoretically or numerically.
(3) Sorry, I really don't have any idea of what "model optimization" is ......
Finally, I recommend you read the book that I mentioned. There are many good examples to warm you up for tackling practical problems by Bayesian statistics.