As far as I know, Ben-Sasson, Chiesa, and Spooner proved in 2016 that for arbitrary interactive protocols with negligible soundness error, the non-interactive protocol generated by Fiat-Shamir transformation also has negligible soundness error under the random oracle model if the original protocol has soundness against state-restoration attacks.
Also, Holmgren showed in 2019 that soundness against state-restoration attacks is equivalent to round-by-round soundness.
I remember that Canetti et al in 2018 claimed that any interactive protocol can be converted into another interactive protocol (parallel repetition of the original protocol) with round-by-round soundness.
Below is my understanding.
From the above results, arbitrary interactive protocols can be securely converted into non-interactive ones under the random oracle model because we can convert the protocol to the one with round-by-round soundness and because round-by-round soundness is equivalent to the soundness against state-restoration attacks, it implies that the protocol can be converted into non-interactive one by using Fiat-Shamir transformation.
Is my understanding true ? I hope that there is somebody confirming my understanding.