I tend to prefer the latter, but I totally get that feeling where signals seem like they "should" be better. I just find in practice that references are a little easier to work with in some cases.

You can easily solve the owner dying issue by just using is_instance_valid() before attempting to call anything on the owner reference.

That said, you can probably simplify your signal code if you connected the bullet killed_enemy signal directly to the player's on-hit handler. It seems like the weapon on-kill handler is redundant? But I don't know the details of your implementation, I just know that there's often ways to simplify chains like this.

I find that signals are great in some use cases, and less good in other use cases.

Definitely the latter. Rube goldberg machines will become much more error prone. The more steps you can simplify out, the better.

You can write a simple guard in case player has died:

if !is_instance_valid(player):
    return

player.onkill()

I have little experience with Godot, so I'll just write the general concept and you'll have to translate.

I'd go with #2, but also have the bullet subscribe to an event that is triggered when the shooter player dies. When that happens, set a private variable that tells the bullet that it shouldn't trigger any on-kill effects (or you could do other things like detonate or remove the bullet).

Something like that in C# (or close to C#, it's been a while):

class Bullet {
    private Player _player = null;
    private bool _enableOnDeathEffect = true;

    private void OnPlayerDeath(object sender, EventArgs e) {
        _enableOnDeathEffect = false;
    }

    public void OnKill(){
        if (_enableOnDeathEffect) {
            //trigger effect
        }
    }

    public Bullet(player) {
        _player = player;
        player.OnDeath += OnPlayerDeath;
    }
}