As part of 94f3b9bfab the code was changed to support dynamically adding localhost to the inventory. This change introduced an crash when run via ansible-pull
```
Starting ansible-pull at 2014-01-20 23:09:57
Traceback (most recent call last):
File "/tmp/ansible/bin/ansible", line 157, in <module>
(runner, results) = cli.run(options, args)
File "/tmp/ansible/bin/ansible", line 82, in run
hosts = inventory_manager.list_hosts(pattern)
File "/tmp/ansible/lib/ansible/inventory/__init__.py", line 372, in list_hosts
result = [ h.name for h in self.get_hosts(pattern) ]
File "/tmp/ansible/lib/ansible/inventory/__init__.py", line 136, in get_hosts
subset = self._get_hosts(self._subset)
File "/tmp/ansible/lib/ansible/inventory/__init__.py", line 177, in _get_hosts
that = self.__get_hosts(p)
File "/tmp/ansible/lib/ansible/inventory/__init__.py", line 198, in __get_hosts
hpat = self._hosts_in_unenumerated_pattern(name)
File "/tmp/ansible/lib/ansible/inventory/__init__.py", line 275, in _hosts_in_unenumerated_pattern
ungrouped.add_host(new_host)
AttributeError: 'NoneType' object has no attribute 'add_host'
```
The root cause is there is no group for the host to be added to. I fixed this case by creating the ungrouped group when it doesn't exist and then adding the host to the newly added group. This fixes the regression for me.
Operate on that play attribute to make things faster for larger
inventories. Instead of making a round trip through inventory.list_hosts
and working through some lengthy list comprehensions over and over
again, calculate the potenital hosts for a play once, then reduce from
it the unavailable hosts when necessary.
Also moves how the %fail is done. The host count is a play level count
of available hosts, which then is compared after each task to the
current number of available hosts for the play. This used to get a new
count every task which was also time expensive.
When applying precedence ordering of different classes of vars
(hostvars, groupvars, role-defaults, etc.), the hash_behaviour
setting controls whether duplicate hash keys are replaced in
entirety, or merged together.
The wording of the documentation suggests that this setting applies
to all levels of the precedence ordering, when it currently does not:
> Ansible by default will override variables in specific precedence orders,
> as described in Variables. When a variable of higher precedence wins,
> it will replace the other value. ... Some users prefer that variables that
> are hashes (aka ‘dictionaries’ in Python terms) are merged together. This
> setting is called ‘merge’.
This change causes the hash_behavior setting to extend to vars plugins.
