[ ex-Intel, although not connected to docs, and I did wind up submitting a number of corrections which got taken up ]
I think "unbelievably bad" is unfair; specifically, compared to what? If you think Intel docs are bad you must not have read many equivalent docs for other processors. We worked on a range of processors for the Hyperscan project before version 4.x, and while Intel docs leave a lot to be desired, you should see the other guys.
That said, there are a lot of things that need to be fixed.
L/T numbers are always idiosyncratic as there are a bunch of ways of getting them and none perfectly predict behavior for all instructions and all cases.
L/T numbers in the ORM frequently 'go missing' which isn't great.
The ORM is a giant 'additive song' that gets a new chapter with each new extension, and tries to provide optimization direction for over 10 years worth of processors. There's still Pentium 4 material in there!
There's somewhat of a culture of secrecy about the architecture that Intel would actually benefit from relaxing. It's clear that other vendors know how to get close to IA in performance or even pass it on some metrics. So I think it would benefit Intel a great deal to help people build better mental models of how their processors work.