Just consistency and performance.

The time goes on and we are a year older already.

false and NULL; typo

what the function actually does.  The previous one was incorrect and just
confusing.

constify things; use DISABLE_ instead of comment out for the disabled test;
typo

assumption on the descriptors; added blank lines for readability.

The test passes on Linux with epoll, as the (invalid) file descriptor is
inserted right away. The others don't detect the error right at the
time, so it doesn't throw.

* Use the address family instead of v6 boolean.
* Check it by exception.
* Documentation.

(based on the seeming intent), indentation.

it without them.

* Use IPv6 address for the test
* Don't use static strings, use char*
* Don't call them static if they are in anonymous namespace
* Use getaddrinfo to get the (slightly system dependant) address structure
* Use protocol field (even if it probably doesn't matter for SOCK_STREAM
  and SOCK_DGRAM)
* Be more careful about releasing resources in error states (even if
  it's a test and they would terminate on the error anyway)

The same problem as with the auth ones - the socket requestor needs to
be faked.

We need to fake the SocketRequestor in them. This also asked for little
update of the general base for socket requesting tests.

* It'll be needed in other tests as well, putting it to testutils.
* Needs documentation.
* A header only, to avoid linking order problems.

* The test fails
* It should at last return and get to some consistent state (eg. not
  listening on anything).

* They are requested and installed, instead of binding them locally
* Documentation needs update
* It'll fail in practice, since the requestor isn't initialized from
  auth and resolver.
* It needs to set the correct IPv6 flag

Only the test initialization and the singleton accessor functions are
implemented for now (and tests for those).

No need to create an opened socket and close it right away.

The dns_service is not tested, so the tests are not added. It is simple
anyway.

reviewed on jabber

details.  This should solve failure of some AXFR tests using SQLite3 on
some systems.