In the cryptography lecture at my university, we had the theorem that (randomized) OFB mode is IND-CPA secure if the used pseudorandom function (PRF) is IND-PRF secure.
Afterwards, we investigated the security of rOFB if a pseudorandom permutation (PRP) is used instead of an PRF. As we can only distinguish a PRP from a PRF if the PRF produces a collision, we can use them interchangeably as long as the block length is large enough. This is the case as for large block lengths the probability of a collision is negligible.
From this result we deduced that we can also use PRPs in rOFB mode which results in a IND-CPA secure encryption scheme.
I totally understand that we can distinguish a PRP from a random function oracle (RFO) if the RFO produces a collision. Therefore, the PRP 'does not look random' if we can query many outputs. However, the security of rOFB mode requires that we do not have a collision in the ciphertexts. Therefore, I do not understand why the security of rOFB mode is lower when we use a PRP instead of a PRF (negligibly lower, but lower) as a PRP does rule out exactly this attack on OFB.
It is clear for me that the PRP-OFB mode does not look like an OTP for many queries, but I cannot imagine why the attack surface should be negligibly higher if there is NO collision possible.