Does rebel use random iteration sampling or subgame resolving?

[thomasahle]

In the ReBeL paper it is mentioned that you use CFR-D modified to stop at a random iteration rather than averaging the strategies; and that this is crucial to not be exploitable. But I don't see the random iteration sampling in subgame_solving.cc. Also CFR-D normally uses a gadget/modified game with "opt out" actions. I can't tell from the paper if this is included in ReBeL, but it seems from the code that it's not?

Does this mean that the solver in subgame_solving.cc is not actually using safe search?