So this isn't necessarily what you'd call a randomiser — often the mods aren't randomly generated at all, but carefully curated by the event organisers. But it's à propos in terms of 'new experience every time'.
One neat aspect of this approach is how unexpected the mods can get; e.g. your player character could be swapped out for a character from a different game. Then there's the extreme variant of mystery races, where the runners don't even know which games they're going to play.
1) There's lots of different categories to compete in. As new possibilities are discovered in the game, the community congrats around goals of speedrunning any given seed. In these scenarios, its generally about building the skills to be proficient at a particular goal in record time regardless of the seed.
2) most of the modern roguelites I play include a "Daily Seed," where every player gets to play the game exactly once each day and be ranked on a leaderboard. You compete on a given seed. If you want a global ranking system, you might consider it by track-record. Something like, "of the last 50 games this player played, what is their aggregate ranking vs other players' last 50 games played", or do that same comparison, but only in the cases where those players shared a seed for the RNG.
It's true that there's no absolute "best game", but there is "best game ever on this particularly interesting seed."
This leads me to believe that this approach can both make speedrunning continue to be new and interesting to participants as well as result in a scenario where the record can mean something. Yes, there are ways to cheat the system if you really want to cheat, in all but a live in-person audited game. It shifts the focus slightly from reverse-engineering of the game and towards being a better all-round player. And it's more interesting for machine solving, too: how can I build an AI that will take the optimal route regardless of seed? How can I know that the optimal route was achieved?