Dynamic class breaks CSS selector

[jesseskinner]

Describe the bug
Some style rules aren't being applied when there is a dynamic class between the root node and target node of the selector.

Logs
N/A

To Reproduce
https://svelte.dev/repl/c122185d662a4897a4cda657d07a3027?version=3.12.0

Expected behavior
I would expect all three lines of the REPL to have red text. Only the second and third do.

Stacktraces
N/A

Information about your Svelte project:
Firefox 69.0
Arch Linux
Svelte 3.12.0

Severity
It's annoying and very unexpected, but I can work around it.

Additional context
N/A

[hbendev]

The problem here is that, the container div doesn't get the svelte-generated class name. The child's class can be any variable declared in a script tag too, and an {#if} tag also causes the same problem, so I guess it can be reproduced with any svelte-tag.

<div class="root">
	{#if true}
		<div class="{aNumber}">
			<div>a number</div>
		</div>
	{/if}
</div>

results in
<div class="root"><div class="24"><div>a number</div></div></div>

[Cleecanth]

Not a fix, but a "simple" workaround is to add an empty .root{} declaration to your style block. This forces Svelte to add a generated class name to the first div.

Thankfully, the blank selector should be easily removed by a minifier while still maintaining the generated class.

[jesseskinner]

As you can see I've added a test for the issue in my fork. I'm very interested in trying to fix this issue, if it's okay with everyone?

I can identify a contradiction in the current code (Selector.ts) that will require a significant change to how encapsulation is calculated in order to fix this issue:

1. A maximum of two nodes (ie. the first and last) are encapsulated per Selector
2. When possible, only the outermost matching nodes of a selector are encapsulated, so we can usually omit the svelte-xxxxx class from the inner matching nodes of a selector.
3. More than two nodes could require encapsulation for a single Selector

I plan to do some refactoring of apply_selector and add additional tests to make this possible without breaking anything. The cheap way out would be to add the svelte class in more cases, but my goal is to do better than this and maintain the existing behaviour when possible.