From that particular post, this line stuck out to me:
Only 3 candidates were able to pass this code filter. The people that did pass either did well enough to pass or extremely excellent. The people that failed, maybe 19 of 22 people, all failed epically.
In practice I highly doubt that the population of candidates is actually this bimodal. It might lean heavily below the hiring bar, but I have trouble believing that it's bimodal.
My own personal experience has been that it looks a lot like a normal distribution where the hiring bar is just two standard deviations above the mean (which would imply that you give offers to about 1 out of 25 people who apply, which sounds approximately in the right ballpark).
As a consequence, if your question is producing bimodal results, I can't help but wonder whether it's a good interview question, because knowledge tests exhibit much of the same type of distribution of outcomes.
In my experience, if you give the candidate only one hour, access to documentation only helps with recalling past experience, because one hour is too much time pressure to actually calm down, pick through, and digest documentation on something you haven't seen before. If you knew the DOM functions existed and just needed their names, it'll work, but if you're unfamiliar with what the DOM already offers, even if you could learn it fine on the job, you won't learn it fine under time pressure. I recently adjusted one of my interview questions specifically to reduce the need for looking up documentation because I noticed exactly this phenomenon.