Ah, I did misread your comment. What you describe is conceptually similar to Svelte's approach in the client, or even signals; keeping a reference directly to the node that used a certain value, though the client-side libraries have the luxury of keeping a pointer to the actual node.

With DOM state being thrown away, it would still not be possible to build as-you-type input validation for example. For SSR + streaming server updates I get the feeling it would also have limited utility, how do you track dependencies across more complex template conditions, loops? Is querying for comment markers any faster than traversing DOM elements? If using generated IDs, do you keep node IDs in memory for each user session in the server when dealing with dynamic content? Are you using an existing open source solution for this?

The DOM diffing/morphing approach is popular because it's in fact extremely fast to run, has small memory requirements, and is a low complexity implementation. In the SSR case, you don't need anything special on the server side, it can be completely ignorant of what is happening in the client. It's hard to beat.