Ed25519 uses a composite order Elliptic Curve but works in the prime order subgroup of the main group. The signing operation should be using a generator/Base Point from the prime order subgroup.
This blog seems to say the following - Honest Ed25519 signers only work in the prime-order subgroup (i.e they chose a generator/Base point from the prime order subgroup). But nothing stops a signer from deviating from the protocol, and generating a verification key with nonzero torsion component (by using input from the cofactor subgroup). In this case, batched verification will succeed but unbatched verification may fail.
I am unable to understand why there is a need to protect against a dishonest signer? Doesn't the whole signing & verification process work on the basis of trusting the signer?
Also, what would the dishonest signer achieve by doing the above dishonesty?