The only failure was exposing public key of a known UTXO address.

Since the public key might as well have been obtained from an already replaced TX, instead of the first one, it's a very weak argument to call it "stealing".Public keys can't be stolen, as they are, as they are called, public information, meant to be known to everyone.

From a solver's perspective, it's just having the additional information (the public key, not the private key). So there is an actual solving effort being done, which is the objective of the puzzle.

Like today's experiment shown, a lot of TXs can simply never make it to a mempool history service, such as mempool.space. Hence, there may always exist a very simple scenario, which nobody mentions:

The first solver's TX may not even make it to mempool.space, because it can very well first be seen by some other node, observed by somebody, replaced, and broadcasted again before the original TX makes it to mempool.space - making the initial TX inexistent in history, as it will be refused when relayed back from the original node (if that even reaches the point of happening).