You can close this ticket. It's a jail-specific behavior when the jail is entered from the host via command line (https://lists.freebsd.org/archives/freebsd-hackers/2022-February/000832.html for the curious people) and something is used which closes all FDs. When the jail is accessed via ssh, it works.

Yes, this is in a jail. But the output above was from the same shell session inside the jail. So gpg-agent was forked from gpg which I executed in the same shell (same process) as the ls. As you can see from the output of ls, /dev/pts is mounted there. The link you provided tells to mount the devfs inside the jail. This is the case here (that's basics, it needs to be there for a lot of things to work inside a jail).

% export GPG_TTY=$(tty)

As the above commit only references pinentry-tty.c, what's the problem with pinentry-curses? Shall I provide the same log with pinentry-curses?

Yes, this was the correct tty at the time of the generation of this log.

I killed gpg-agent after the config change / before running gpg again. That should be enough to pick-up the config change, correct? In the mean time the system in question was rebooted. Here the full log /with key related stuff redacted).

GPG_TTY in my tests from which I generated the logs above is set to /dev/pts/1 (which is what "tty" returns, what exists in the FS, and what is writable to the user which performs the test).