Fix x-mask attribute changes on x-model inputs

[joshhanley]

The Scenario

When Livewire morphs an input that uses both wire:model and x-mask, changing the server-rendered x-mask attribute can leave the input using a previous mask.

In this example, the input starts with a five-digit French postal code mask, changes to a four-digit Belgian mask, then changes back to the five-digit French mask. The x-mask attribute updates, but the input remains limited by the older four-digit mask.

use Livewire\Attributes\Computed;
use Livewire\Component;

new class extends Component {
    public string $country = 'FR';
    public string $postalCode = '12345';

    #[Computed]
    public function mask()
    {
        return $this->country === 'FR' ? '99999' : '9999';
    }
};

<form>
    <select wire:model.live="country">
        <option value="FR">France</option>
        <option value="BE">Belgium</option>
    </select>

    <input wire:model="postalCode" x-mask="{{ $this->mask }}">
</form>

The Problem

The mask directive wraps el._x_forceModelUpdate so programmatic model updates are formatted before they are applied to the input.

When an existing x-mask attribute changes, Alpine treats that as directive cleanup followed by directive initialisation. The cleanup aborts the directive's event listeners, but it does not restore el._x_forceModelUpdate.

That means every reinitialised mask wraps the updater currently on the element. Previous masks stay in the update chain, so an older four-digit mask can still trim a value after the element has changed back to a five-digit mask.

The Solution

The solution is to let the mask directive unregister the model updater wrapper it installs.

The directive now keeps a reference to its wrapper and restores the previous updater during cleanup, but only when the element is still using that exact wrapper. This avoids removing a newer wrapper that may have been installed after another directive initialisation.

Fixes livewire/livewire#10377

[calebporzio]

Traced this through mutation.js (removed-attrs cleanup runs before added-attrs init, so the wrapper chain stays depth-1) and verified the Cypress test fails on main and passes with the fix. The identity guard on restore handles the x-model-reinit case nicely, and dropping the cleanedUp microtask guard is safe since any dead wrapper gets clobbered by x-model re-init. Ready to merge.