Reactivity degradation in 5.35.5

[gyzerok]

Describe the bug

Here is the PR that causes the problem it seems: #16278

In our codebase we are using mutable deriveds for optimistic updates. Whenever there is a list to be updated we use SvelteSet or SvelteMap. In a certain cases we are calling update in onDestroy because we want to actually perform it whenever context menu is closing. Now in 5.35.5 it got broken.

Unfortunately we really needed another PR that got merged within 5.35.5, so now we are blocked from upgrading.

Reproduction

REPL

1. Click on the button
2. In 5.35.4 number 2 will disappear from the list
3. In 5.35.5 it does not

The repro could be trimmed down further, I just wanted to avoid questions "why are you doing this". Hopefully it's a bit clearer this way.

Logs

System Info

no relevant

Severity

blocking an upgrade

[gyzerok]

Funny enough this could be fixed by wrapping delete in setTimeout like so:

onDestroy(() => {
  setTimeout(() => {
    set.delete(2)
  })
})

[PatrickG]

This is a bit strange.
It feels like test.ids is a different Set instance inside of the callback function.

But the setTimeout workaround (un/comment line 7/8 in Teardown.svelte) is strange as well.
Now test is the correct Set instance inside of the callback function.
But after clicking "click", test.ids in (my added) append function seems to be a different Set instance.

https://svelte.dev/playground/3e12940167f54f7db44f62e634a28499?version=5.35.5

[Rich-Harris]

It is a different set — what's happening here is that when test.ids is read in the callback, Svelte wrongly believes that it needs to re-run the derived, because (I think) this line is wrong:

svelte/packages/svelte/src/internal/client/runtime.js

Lines 656 to 659 in b17557c

	// if the derived is dirty, or depends on the values that just changed, re-execute
	if ((derived.f & CLEAN) !== 0 || depends_on_old_values(derived)) {
	value = execute_derived(derived);
	}

That !== 0 should be an === 0, and I'm pretty sure the second half shouldn't be necessary at all. Unfortunately that causes some other tests to fail, because of reasons I haven't quite diagnosed yet.