Fix really_destroy used against a has_one relationship that is soft deleted

[reedstonefood]

I found a bug. The test explains the scenario but I'll explain it here as well.

Given:

class MyModel < ActiveRecord::Base
  acts_as_paranoid
  has_one: linked_model, dependant: destroy
end

class LinkedModel < ActiveRecord::Base
  acts_as_paranoid
end

my_model = MyModel.create!(linked_model: LinkedModel.new)

• ✅ If you run my_model.destroy then both objects get soft-deleted as you would expect
• ❌ However if you then subsequently run my_model.really_destroy it does not destroy the associated linked_model.

This is because paranoia tries to access the linked model by running self.send(name). This command will not pick up soft-deleted has_one associations (though it works fine for has_many associations) so some extra code is required in this scenario.

Note the reload in the test is required, as otherwise the version of my_model in memory still links to the LinkedModel.

[reedstonefood]

@mathieujobin any chance you could have a look at this please?

[reedstonefood]

FYI the test failures on rails versions older than 7.1 are nothing to do with the changes in this MR, they are a wider issue that is detailed here: https://stackoverflow.com/questions/79360526/uninitialized-constant-activesupportloggerthreadsafelevellogger-nameerror

[reedstonefood]

I've just spotted that this looks like a duplicate of #428, but as I'm actively engaged about this MR and it has working tests, it probably makes sense to progress with this one.

[mathieujobin]

@yoheimuta Can you chime in about this change?

[yoheimuta]

@mathieujobin Looks good to me! 👍

[nits] For has_one associations that are polymorphic (foreign_type column exists), find_by should also filter by foreign_type to avoid fetching unrelated records.

Suggested improvement:

query = { reflection.foreign_key => self.id }
query[reflection.type] = self.class.name if reflection.options[:as] # Handle polymorphic associations
association_data = reflection.klass.only_deleted.find_by(query)

This ensures we correctly match records when dealing with polymorphic has_one associations.

[mathieujobin]

@reedstonefood Can you review this suggestion, see if that matches with your use case, etc.

thanks

[reedstonefood]

This change works with the test suite here & in the product I'm working on, so LGTM.

I've also managed to add another test that fails without this change in place 👍 .

[mathieujobin]

is it possible to fix the tests for older version in this PR or in a different PR ?

[reedstonefood]

is it possible to fix the tests for older version in this PR or in a different PR ?

I've raised #576 for this.