One more trick is to lower the highest ionic frequencies by turning the protons, that are not crucial to the process you want to investigate, into deuterium. Their bonded vibrations may be of pretty high frequency and thus slowing them down by increasing their mass may help improving the stability w.r.t. timestep and fictitious mass.
Summarizing, it might happen that you will not be able to assure perfect adiabaticity. Nevertheless you always have to make a choice between numerical accuracy, algorithmical correctness and statistical convergence. It is very important to quantify the amount of error that you are going to introduce, but then you have to weight it against the systematic error of the method (DFT, pseudopotentials, CP-dynamics, ergodicity) and make a balanced choice. For instance the typical error for DFT calculations can be easily in the 1-2 kcal/mol regime, e.g. due to the exchange-correlation functional used. If the energy drift rate has the order of 0.05 kcal/mol per picosecond and what you can afford is 10 ps simulation, the total error introduced by non-adiabaticity is 0.5 kcal/mol during the whole run. Since this seems tolerable although not perfect, you can consider using a thermostat on the electrons during you production run.