Is there any evidence that candidates who do well on design interviews but fumble on leetcode would be any worse at the actual job?
people who cannot leetcode - a simple data structures & algorithms inteview - cannot understand runtime nuances of what they write.
This is how you end up with N+1 algorithms and exponential runtime.
For example look at all modern front-end in javascript - usually leetcode is more relaxed for front end, and you end up with ungodly gobbles of unnecessary loops inside loops wrapped in loops that traverse DOM for no good reason and freeze the browser.
it is the javascript people who import node libraries that contain 3 lines of code, instead of writing it properly
Primed recall can remain strong for many years, but leetcode often penalizes googling around to refresh your memory, even though that is what almost everyone should do before taking any further steps toward an implementation. Depending, it's often also the correct thing to do before choosing a library, the parameters to a function, code base organization, and lots of other details.
Of course, domain experts may be very strong in their specialty, but that's a special case with narrower scope.
I can hypothesize that testing for a "fast study" may have more predictive power. I have seen some interviews that are designed for this, leetcode adjacent but less antagonistic.
But anyway, I am spitballing, to be real with you.