Judging from the odd Torvalds-yelling-at-people mail getting linked on HN, it doesn't sound like "it's ok to break userland here because clearly they didn't read the docs" would fly.
In the mental model in which the /dev/random manpage lives, it makes (some) sense to read /dev/random until you get a short read, then make a userland pool from all the data you've read, "to get as much entropy as possible". That reasoning doesn't actually make sense, of course, but I would be surprised if nobody thought that such a construction would be a good idea.
Separately, there's quite a lot of code - and I've written some - that reads a byte from /dev/random at daemon start, specifically to block until the system has built up enough entropy. FreeBSD blocks reads, even from /dev/urandom, until the [EDIT: estimate of how much entropy has ever gone into the pool] gets high enough, which makes some sense; but there needs to be some way to block-until-random.
I'm fairly sure they would have to change the nature of /dev/urandom in such a way as would invalidate the other parts of the documentation in order to make the erroneous parts of the man page contain valid concerns.
Just because the implementation of /dev/urandom has good properties that make usable as a system-wide CSPRNG, doesn't mean that it's supposed to have those good properties. Only the docs can make that assertion. Without such an assertion, the kernel is free to make /dev/urandom have less-sound properties in a future release—because there's no spec saying they shouldn't. (They try not to break userland code, yes, but userland code only relies on things the docs say it can, so...)