It is right if you are only interested in making the damn thing work.

But if you don't want it to broadcast noise like crazy you want it tight and matched as much as possible.

The value of differential pair comes from the fact that they have constant common component. That only works if both components aren't shifted in time. If that happens, you will get strong common signal and say bye to nice quiet differential pair.

Are you saying your differential line is causing EMI if it's not length matched?

Or are you saying that the receiver won't do a good enough job cancelling common mode noise if you're not length matched? For that, I'd like to have a reference. Otherwise I'm not convinced this is more critical than the max skew to keep your crossings in the right region.

Differential pair minimises EMI by having constant common mode potential over the length of the line (and wires close to each other). Meaning if you look at the pair from afar the average of their voltage at any point in the line is constant. Or you might say that the noise from both lines cancels each other. From afar, it looks like there is no signal traveling the length of the line (to an approximation, because there is necessarily some distance between the wires meaning the cancelling is not perfect).

If the signals are shifted in time, though, that is no longer true. From afar you will see places in the line that have different common mode potential. Where one signal edge lags behind the other the average will no longer be the same as for the rest of the line. From afar this looks like high bandwidth signal travelling down the line.

I mean ... he's saying right there on the slide there's less than 1mm of length error budget at 10GHz.

And that's gobs. What's your typical space between the lines in a pair? Five mils? You run them side by side and you'd be hard pressed to make a length error significant enough to matter. Yeah you do need to be aware of it enough to ask whether you're good when you're routing a really weird loop.. but for the vast majority of things and sane routing, you can just eyeball it or don't need to care at all.

Point is, a lot of people seem to think length matching is critical ("for anybody that is working on anything over couple hundred megahertz") and application notes would have you believe as much but the industry experts and theory I've found suggest it's way overstated.

I've always followed Howard Johnson's rule of thumb to match them within 1/10 of the rise time, but it's not like it's a lot of work: the software does it automatically.