fwiw, when i'm on a network that doesn't support IPv6, i get this:

0 dkg@alice:~$ gpg --send $KEYID
gpg: sending key REDACTED to hkps://hkps.pool.sks-keyservers.net
gpg: keyserver send failed: Invalid argument
gpg: keyserver send failed: Invalid argument
2 dkg@alice:~$

in dirmngr's logs:

2016-06-17 19:30:17 dirmngr[27999.2] DBG: gnutls:L3: ASSERT: mpi.c:246
2016-06-17 19:30:17 dirmngr[27999.2] DBG: gnutls:L5: REC[0x7f61f400fc10]:
Allocating epoch #0
2016-06-17 19:30:17 dirmngr[27999.2] can't connect to '2001:ba8:1f1:f2d4::2':
Invalid argument
2016-06-17 19:30:17 dirmngr[27999.2] error connecting to
'https://[2001:ba8:1f1:f2d4::2]:443': Invalid argument
2016-06-17 19:30:17 dirmngr[27999.2] DBG: gnutls:L5: REC[0x7f61f400fc10]: Start
of epoch cleanup
2016-06-17 19:30:17 dirmngr[27999.2] DBG: gnutls:L5: REC[0x7f61f400fc10]: End of
epoch cleanup

I think this instance of dirmngr was started on a network that has both IPv4 and
IPv6.

if i do:

     gpg-connect-agent --dirmngr killdirmngr /bye

and then try the --send again, it goes through fine.

(that last comment was with 2.1.13)

I think that this is same issue of T2438: dirmngr fails repeatedly with "invalid argument", without kicking the host from its list.
Merging.

I tested with:

# sysctl net.ipv6.conf.all.disable_ipv6=1

Now (2.2.8), it (connect(2)) returns EADDRNOTAVAIL, which is correctly handled by handle_send_request_error.

As written in T2438:

One problem was fixed for handling IPv6 name resolution: rG892b33bb2c57: dirmngr: Fix alignment of ADDR.