Hi mmalmi,
If the object of interest for you is the stochastic (ergodic) mean of welfare, it might be the case that you are not interested in the graphical visualisation of IRFs, as most often it is the case. Therefore, you only need to solve the model at second-order and switch IRFs off. For ergodicity, you also need many replications, say 1000.
If you are interested in the graphical visualisation of IRFs, then pruning and reducing the size of the shock are both viable options, provided that one bears in mind that shock size matters in a second-order approximation due to the absence of certainty equivalence. Personally, I believe reducing the shock size is a slightly poorer substitute for pruning, especially if you have an estimated shock size that you are reluctant to change.
If you produce enough replications, say 1000, the simulated IRFs (higher-order IRFs are called GIRFs because they are state-dependent) will not be lumpy.
Try
stoch_simul(order=2,irf=0, replic=1000);
You can always increase the number of replications to make GIRFs smoother.