empty string, so check_prepare() can't expect factory.prepare() to
change the host on Windows.  This repairs Windows failures in

test_fcgi_factory
test_ftp_factory
test_http_factory
test_icp_factory
test_webdav_source_factory
test_monitor_factory_with_emergency_user
test_monitor_factory_without_emergency_user

Unfortunately, it also exposes 2 new Windows failures in

test_http_factory
test_webdav_source_factory

       'sort_on'.

     - PUT would fail if the created object had a __len__ = 0 (eg:
       BTreeFolder2) and fallback to _default_put_factory. Fix by
       checking if the returned object is None instead.

when an 'expires' keyword argument was passed.

avoiding the weird handler calls makes the initialization easier to understand
(and therefore maintain)

debug handlers installed much earlier (before we have any configuration info)
so that very early log messages are not lost.  The debug handler is removed
as soon as we know whether debug mode is disabled, and the startup messages
are flushed to the real log as soon as we have it.

The Zope.Startup.ZopeStarter object makes more sense now as well.

  - Remove fossil tests.

these must be controllable separately

(the tests themselves still aren't being run; we need decent
save/restore behavior from the logging package to do this properly)

later

See http://zope.org/Collectors/Zope/1289.

Avoid the monkey patch by creating a subclass of Connection and
installing it as the klass attribute of ZODB's DB class.  I think this
works correctly in conjunction with Zope configuration (the tests
pass) -- but I'm not qualified to say for sure.

Rework the logic of MountConnection as part of the restructuring.  The
_root_connection attribute always points to the root, never to None;
so the root object points to itself.  This change necessitated a
chance in DBTab.ClassFactories.  Add some comments there to explain
what's going on.

XXX The close() method on MountConnection is re-implementing two
methods in modules.  If those implementations change, it would need to
change, too.  We ought to find a better way to integrate it.

XXX Perhaps the mounted connection stuff should be folded into ZODB
proper.

just like it does not exist

the signal handler installation is abstracted in the same way