Running the model rotdk with GAMS/Cplex yields:
Solution satisfies tolerances. MIP Solution: 2012.831285 (802 iterations, 0 nodes) Final Solve: 2008.641285 (0 iterations) Best possible: 1921.405150 Absolute gap: 91.426136 Relative gap: 0.045422
If a MIP has continuous variables, the objective of the fixed model can differ from objective value reported when the incumbent solution was first found. This can happen because CPLEX has many different procedures to derive integer feasible solutions. Some of them provide an optimal basis to a node LP that will also be optimal for the fixed problem, but others do not. In the latter case, changes to the continuous variables in the fixed model may yield an improvement in the objective and this is what is happening with the model. (reference).
This behavior will disappear, if you set optca/optcr to Zero.