I expect you’d get a better result by picking a radius such that you get slightly more samples than you need, and then removing a few.

Is the spelling "Poission" intentional?

Oof, thanks

Maybe I'm misunderstanding, but it seems like the points placed in the second pass aren't checked against the Poisson disc condition. So should there be a disclaimer that says that the result is actually not a conforming distribution?!

Good call. I've added the disclaimer. Thanks.

I implemented another approach in https://observablehq.com/@fil/poisson-finish — probably a bit similar to what Jacob had in mind. I can now close this browser tab :)