I've read that paper so many times before in so many ways. What I can say on the matter is that the paper does not describe explicitly what the framework particularly does. It just gives an hint of their intuitive idea of bundling of the features in an efficient way. But specificly, it does not say that it does a 'reversion of one-hot-encoding' in particular to your question.
I tried giving categorical inputs directly and as one-hot-encoded to compare the time that it takes to compute. There was a significant difference: giving directly was all better in multiple datasets compared to giving as one-hot-encoded.
1)It is possible that LightGBM Framework can find out that we give the features as one-hot-encoded from the sparsity, it is possible that the algorithm does not treat one-hot-encoded with EFB.
2)It is also possible that LightGBM uses EFB on one-hot-encoded samples but it may be harmful, or not good as EFB on direct categorical inputs. (I go for this one)
But still, I do not think that EFB will reverse one-hot-encoding since EFB is explained as a unique way of treating the categorical features. But it possibly 'bundles the unbundled features' when treating one-hot-encoded inputs.
I used the word 'probably' so much times out of implicitness of the paper. What I can advice to you is that to send an e-mail to one of the authors of the paper, I do not think that they would refuse to explain it. Or if you are brave, go for the GitHub Repo of LightGBM, to check the codes by yourself.
I hope that I could give you an insight. If you come up with an exact answer on the matter, please let me know. Please do not hesitate to further discuss this, I'll be around. Good luck, have fun!