This was broken in Rails 4.2.0+ because the `@scope` object is no longer a Hash
but an internal structure that supports a better override/rollback flow for cases
like this. If we would only support Rails 4.2, this method could be something
like this:
```ruby
def with_devise_exclusive_scope(new_path, new_as, options)
overrides = { as: new_as, path: new_path, module: nil }
overrides.merge!(options.slice(:constraints, :defaults, :options))
@scope = @scope.new(overrides)
yield
ensure
@scope = @scope.parent
end
```
I was seeing the following in my console:
```
/home/vagrant/.rbenv/versions/2.1.2/lib/ruby/gems/2.1.0/gems/devise-3.2.4/app/controllers/devise_controller.rb:9: warning: `*' interpreted as argument prefix
/home/vagrant/.rbenv/versions/2.1.2/lib/ruby/gems/2.1.0/gems/devise-3.2.4/app/controllers/devise_controller.rb:10: warning: `*' interpreted as argument prefix
```
This change silences this warning.
This leaks into Devise mappings overriding the existing :admin one, so
depending on the test seed it fails like this:
$ TESTOPTS="--seed=9972" rake
==> Devise.orm = :active_record
Run options: --seed=9972
...
1) Failure:
MappingTest#test_allows_path_to_be_given [./test/mapping_test.rb:31]:
Expected: "admin_area"
Actual: "admin"
Using a different name should avoid any test randomization issues.
Previously the test was raising an ArgumentError by mistake:
ArgumentError: wrong number of arguments (0 for 1)
actionpack (4.1.4) lib/action_dispatch/routing/route_set.rb:328:in `eval_block'`
The #eval_block method expects a proc/lambda argument that can be
instance_exec'ed, not a real block. In the Rails implementation the block
is passed to #draw, which calls #eval_block internally passing the block
along, but as a Proc argument and not as a block.
Also the error we were raising from #devise_for was a RuntimeError,
changed to the expected ArgumentError. Adding an assertion on top of the
expected message should ensure we won't have this issue again.
This was introduced in 29da146c07, related
to #2802.
Also refactor tests to remove the custom failure app class only used
once in favor of an inline class for the specific test, makes it easier
to follow what's going on.
