0. sshd will fork() and chroot() into /var/empty. After the fork(), you can't use the entropy you have because it's shared with the parent (i.e., it's not "entropic").
1. Where should it get entropy from?
2. Where does OpenSSL get entropy from in this case?
Anyway, I don't even work on portable libressl because I don't want to deal with shit like this, but I think the quoted text erroneously gives the impression that /dev/urandom isn't used. I wanted to correct that impression.
I really think that LibreSSL's RAND_poll() should have similar behavior to ensure maximum compatibility with OpenSSL and to provide a means to use chroot() safely without the risk of falling back to the "bobo" code. (Also, a means of safely reseeding after a fork on systems without minherit(MAP_INHERIT_ZERO)).
(Incidentally, libsodium has a similar API: you can explicitly call randombytes_stir() to cause the library to open (and keep open) a file descriptor to /dev/urandom, so subsequent calls to the PRNG work in a chroot.)
The presence or need for a stir() function should be considered a serious design flaw.
(forks are detected by calling getpid() if you don't have inheritzero.)
Also, getpid() isn't airtight - if you fork and fork again there's a risk of PID wraparound.
What's wrong with generating some random numbers using the parent's entropy pool before fork and using that as the child's entropy pool?
> An error occurs if the PRNG has not been seeded with enough randomness to ensure an unpredictable byte sequence.
https://www.openssl.org/docs/crypto/RAND_bytes.html
But there is also this:
> Pseudo-random byte sequences generated by RAND_pseudo_bytes() will be unique if they are of sufficient length, but are not necessarily unpredictable.
In the former case they are sacrificing portability for increased confidence of correctness, and in the latter they are sacrificing confidence or correctness for increased portability.