It is probably not out-of-bag (in the ensemble sense). It depends on the model fit since what you see in the fit slot comes from the underlying model object.
I completely recommend using predict() to get what you want.
Thank you for your recommendation. On the other hand, it's just super convenient to use the workflows, but I understand that for anything more specific you have to resort on the more generic/basic building blocks.
Sorry, I forget to mention that I was using the ranger engine.
Lately, I came across to this issue - based on which it should be the OOB prediction (and my results were suggesting the same).