If I'm reading the article correctly, the user can be informed at runtime with little to no performance impact. That certainly sounds preferable to returning an unsorted list (that may not even match the input).
Almost every Rust sort does this, but no C/C++ sorts do. That appears to support that author's conclusion that this is a cultural thing.
It's not possible to verify a function is a strict weak ordering without enumerating the entire domain.
The idea is that the implementation doesn't promise it will detect a strict weak ordering violation, but it has code that will detect the effects of such violations with a high probability. It's not a sampled approach, but rather as part of the small-sort a bi-directional merge is performed and if the pointers don't line up, the only explanation is a strict weak ordering violation, the result of the merge is discarded and a panic is raised.