Importantly, it also allows you to examine fluency. The candidate who can write a code/pseudocode solution as fast as they can type, and explain it verbally at the same time, is much more impressive than the candidate who pauses for a long time before starting haltingly and unconfidently. Fluency in solving leetcode problems just tests for having practiced leetcode, but fluency in writing real code tests for actual productivity.
If you have a question like this, you can drill down into performance characteristics just as the question in TFA does. Not necessarily using big-Oh notation, but "what if you have to do this millions of times? what if you have to provide results immediately? what if some of the files are too big to fit into RAM?". But you can also drill down into other aspects of the problem like changing specs, noisy input, constrained resources, observability.
FWIW, when I came up with a question like this (with the constraint that it had to represent a problem I had solved that morning), ChatGPT was not able to solve it correctly.