I've definitely had the experience of "I think this is a valid proof. I haven't seen it presented anywhere else, and I'm not super familiar with the subject matter. It feels a little socially awkward to ask someone to check if it is right", and I started trying to express it in Lean.
However, I found it difficult to get the implicit type parameters to work right, so I had trouble even formulating the claim in Lean. (the claim was a claim in category theory, and the category theory library for Lean has lots of implicit argument stuff, and probably shouldn't have been the first thing I tried to use Lean for. Also the claim I was trying to show is probably very simple for category theory people.)
A proof system could identify cases where you need hidden assumptions in order to make proofs. In this case, you would get a null result from the oracle, which would be valuable in itself.
The proof checker could detect such errors in a purportedly clean proof.
For the second, understanding why and actually proving something can be related but they're not equivalent. Of course, the intuitive ideas can lead to the right formal technicalities but it is not always the case.
There are two aspects of mathematical proofs-communication of ideas and formal verification. Confusing these distinction can be dangerous and this is why mathematicians led to the area of type theory and formal proofs.
Voedovsky, you can google him to find who he was, lost his faith in "human proof checking" after some error on his peer-reviewed and published result was accidentally discovered after years of nobody doubting its validity.
He dreamed of the world where people do and enjoy mathematics freely leaving the boring parts to computers like mathematicians in nowadays don't worry about how their writing will be published and communicated. What we do is write a proof and leave actual typesetting to the computer. It's called (La)TeX.
What you're saying is like claiming this program is correct because its logic is correct without ever knowing whether the code itself would get compiled or not. More extremely, there's no value for having more and more pseudocodes that we don't know they'll ever terminate or not.
You've got a point, but in this case you are barking at the wrong tree. Aren't they talking about automated proof-checking?