The problem with leetcode is time. If you know a good solution to a puzzle, or can instantly come up with it, you have the time to write it down, make it work, discuss, improve, etc. If you spend extra 10 minutes on a wrong approach, you're very strained, and if you spend 15 minutes, you're toast.

But knowing the good solution by heart is not the point! For that reason, when I ran code interviews, I often gave hints to the candidate, or sometimes just stated that a particular algorithm is a good solution, and it's usually implemented using this and that. ("Use BFS. You'll need a queue.")

This is where the demand for realistic conditions kicks in: at work, you're not going to invent an algorithm, you'll ask the machine what works in a situation like yours, consider, and choose. Doing otherwise would be imprudent.