I started reading this a bit skeptically, but my main concern was addressed: this doesn't reimplement JSON deserialization; eserde's deserializer first uses serde_json's deserializer, and if it succeeds, it returns the result. So you don't have to trust someone else's deserializer that might not be maintained as well as serde_json is.

The downside to this approach is that if serde_json's deserializer encounters an error, they then have to deserialize the input again with their own deserializer (that wraps serde_json's) in order to get all the errors out. Fortunately the happy path is just as fast as using serde_json directly, but the unhappy path will be at least twice as slow as serde_json when it encounters an error. But I think this could be an acceptable trade off for some people.

What would be really nice, though, would be to change serde/serde_json/serde_${WHATEVER} so that it can do this (optionally, since even with built-in support, the failure case will still be slower and take up more resources). I expect it could be done in such a way to maintain API/ABI compat with the current state of things.