Performance benchmarks:

As _HardKeyAgent doesn't inherit from Sequence, DeffuantWeisbuchModel is raising TypeError

        if not isinstance(population, _Sequence):
>           raise TypeError("Population must be a sequence.  "
                            "For dicts or sets, use sorted(d).")
E           TypeError: Population must be a sequence.  For dicts or sets, use sorted(d).

As _HardKeyAgent doesn't inherit from Sequence, DeffuantWeisbuchModel is raising TypeError

We have to fix the model anyway because of the deprecation. The source seems to be self.random.sample(self.agents, 2). Which is easily replaced with self.random.sample(self.agents.tolist(), 2).

I think to_list() is quite good improvement
For eg. In this DeffuantWeisbuchModel

# Looks like O(N), but is actually O(N²)
for _ in range(self.n):
  # random.sample calls __getitem__ repeatedly.
  # Since our __getitem__ rebuilds the list every time (O(N)), this line is expensive.
  agent_a, agent_b = self.random.sample(self.agents, 2)

Initially user was doing this, unaware of what was going under the hood
Now this API explicitly tells them what they are doing giving them chance to optimise

agents_list = self.agents.to_list()
for _ in range(self.n):
        agent_a, agent_b = self.random.sample(agents_list, 2)

# Truly O(N)

However this was only possible because loop didn't involve any addition or deletion but I think it is still
cool

Initially user was doing this, unaware of what was going under the hood

And this was exactly the reason why I wanted to deprecate it. I saw other models doing exactly this without realizing the overhead.

Coverage sits at 87%. Can you check this and see if you can add relevant tests to bring this up? I'll try to review later today.

Can you check this and see if you can add relevant tests to bring this up?

Done

One remaining question regarding how to properly document this. Codewise, I am fine with merging this.

I see two options:

1. Update it in this PR only,
2. We can update documentation in next PR where I am planning to remove the self._agents and updating documentation

If you prefer option 2, I will revert the changes in model.py

I prefer option 2, we can finalize the storage refactor there

Option 2 is fine with me as well. I'll take a last look at this and merge it hopefully this afternoon.

PR title and description updated to reflect reverted model.py changes