I would rather descenders weren't subsumed by the line either, but interrupting the element itself instead seems like as much as a problem as it is a solution.
Like is often the case with design, there are trade-offs. In this case I think the readability improvements to links with descenders outweighs the visual cue / scannability of links, particularly because I think links with more-descenders-than-non are not as common.
An interesting experiment might be to try to turn off the effect on words which fail some check (a high percentage of descender glyphs, for example). I created an issue on GitHub [1]—definitely something worth exploring. Thanks!
As an aside, I love Butterick's link style, what with the triangle and the background-changing color hovers, but that's too complicated for a general audience--especially one reading on the newest touch-screen whatsit that can't show you the "◊Amazon" word is actually a link to Amazon until you accidentally tap it.
But that's a great style for reading long-form articles because your brain isn't constantly interrupting you to say, "Ah, please compulsively hover over that underlined text so I can see what website this insightful blogpost is referencing."
I have a hunch using a faded color for the (possibly-then-thicker) underline would work better - the foreground descenders would retain distinction, without the noise of dotty/dashy line-ends.