Replace SymbolTree with a DOM-specific tree

[scttcper]

This came out of profiling jsdom in my DOM benchmark suite.

Replace SymbolTree with a DOM-specific tree stored in NodeImpl._links. Keep link bookkeeping in the tree helper and expose traversal and tree comparisons as methods on nodes.

Allocate links only when needed, preserve sibling indexes before mutations, and calculate mutation indexes only when live Ranges need them. Walk shadow trees with an explicit stack while preserving behavior when callbacks change the tree.

Limit Range stringification and deletion to the selected content, and return immediately when stringifying a collapsed Range. Reuse common-ancestor results in Range operations. Skip descendants of nodes being deleted, and check the two partially contained children in surroundContents() instead of scanning unrelated content.

A few places we looked for inspiration:

• Gecko’s shadow iterator: use a stack instead of recursive generators.
• Gecko’s child-index cache: keep indexes for siblings before the mutation instead of invalidating them all.
• WebKit’s traversal helpers: walk directly through child, sibling, and parent pointers.
• Blink’s compact sibling links: tried smaller node records. Mutations were slower in JavaScript, so we kept a separate pointer to the last child.

Execution time measured at 39edbb73 versus the SymbolTree base, measured on Node 26.8.2 in two paired runs. Percentages combine both runs; cloning varied enough to show the observed range.

A live Range is a DOM Range whose start and end positions automatically update when the tree changes. Keeping those positions up to date adds work to mutations.

Benchmark	Change
Repeated middle insert/remove, without live Ranges	94% less time
Repeated middle insert/remove, with a live Range	93% less time
Random sibling reordering with a live Range	61% less time
Adopt a 100-field subtree	39% less time
Clone a 100-field subtree	Mixed: 17% less to 5% more time
Shallow clone	7% less time
customElements.upgrade(), flat/deep/field-heavy trees	42–64% less time
Element child/sibling getters, adjacent elements	46–49% less time
Element child/sibling getters, mixed text, comments, and elements	52–56% less time
Element child/sibling getters, no matching elements	63–66% less time
Range common ancestor, deep branches and shared ancestors	92–99.9% less time
Range.deleteContents(), fully contained deep subtree	99.4% less time
Range.surroundContents(), adjacent text with 1,000 unrelated siblings	99% less time

Fixes #4291

[domenic]

This looks very promising. I haven't gone into full depth, but I think we could do better from a code readability perspective. To me this looks too much like the old shape.

For example:

• I don't think we need to use a symbol; we can just use node._links or similar.
• I don't like the internal-constants.js file or the various indirections like threading it through shadow-dom.js.
• At the very least, we should change helper calls from e.g. domSymbolTree.index(childImpl) to treeHelpers.index(childImpl) or treeOps.index(childImpl). But I think it'd be nicer if we could do something like childImpl._treeIndex().

I'd also prefer to avoid adding unit tests of internals in test/api. if we think they're really valuable, we can consider a new test/unit/ or something, but I'm hopeful the coverage isn't needed since WPTs cover everything exhaustively.

[scttcper]

@domenic its a really interesting area! Looking around for inspiration that already exists was pretty neat.

Switched to _links, removed the import indirection, and moved index calculation onto NodeImpl. Removed the internal tests and kept the useful coverage in WPTs.

[domenic]

I set Astra Extra High loose on this and it found more cleanups and performance improvements, especially in ranges.

I'm going to move github.com//pull/4308/commits/1feedc0cd62401ac6aceafbb194497606e840b2a to another PR, along with the same fix for processing instructions, then rebase this on that...

[scttcper]

Also using astra xhigh

I ran the benchmarks before and after your commits on Node 26:

Benchmark	Less time
Common ancestor in a deeply nested tree	99.9%
Delete a fully contained deep subtree	97%
Surround text with 1,000 unrelated siblings	97%

Pushed a follow-up to skip unrelated subtrees in toString() and deleteContents() too.