Add indexed and named NamedNodeMap access

[andrewiggins]

Summary

Add live browser-style property reads to Element.attributes without making NamedNodeMap inherit from Array or proxying each instance:

element.attributes[0];
element.attributes.id;

• Resolve indexed and named reads through one shared proxy in the prototype chain, preserving collection identity and avoiding a per-map proxy allocation.
• Match named attributes by qualified name across namespaces while preserving methods, inherited named properties, and ordinary expando precedence.
• Parse only canonical ECMAScript array-index names ("0" through "4294967294"), rejecting fractions, aliases, and out-of-range values before linked-list index lookup.
• Preserve current mutation and custom-element reaction behavior.
• Add focused regression coverage, WPT classifications, and a minor changeset for @remote-dom/polyfill.

This is the focused property-access portion of #693, rebuilt on current main. Removal/replacement fixes and attached-Attr adoption are not included.

Design and performance

Fresh measurements on 2026-09-17 replace the historical benchmark claims from #693. The original harness/raw samples were unavailable, so these are reconstructed microbenchmarks, not an exact reproduction of that experiment.

Environment: Apple M4 Pro, arm64 macOS 26.6.2; Node 24.19.0 / V8 13.6.233.17-node.51; pnpm 10.28.0. These are Node/V8 measurements, not browser or application-level performance claims.

Compared exact base a13aa618182d5c84d1c733977d1ac66955b95dbc with PR head 2f9c14177749f30b3f18f82ad85c162c3d55a9b3, built with the same pinned toolchain. A third, scratch variant applies the rejected constructor-returned per-instance proxy design to the current head while retaining its mutation logic and strict parser; it does not compare against the older branch's unrelated code.

Runtime

Five fresh processes per variant/workload group, run serially in rotating variant order. Each process used 15 warm-up and 15 measured batches calibrated to roughly 10–15 ms, with observable checksums. Established-operation fixtures rotate through 257 elements with four attributes. Times include fixture selection/checksum overhead. Reported ratios are the median paired head/base time ratio, with the min–max across five process pairs—not confidence intervals. Lower is faster; differences around ±5% are treated cautiously, with no aggregate workload score.

Workload	Head/base time	Five-pair range
getAttribute hit	0.95×	0.88–1.11×
Replace existing attribute, toggling value	0.99×	0.97–0.99×
Remove + re-add attribute (paired operation)	1.02×	0.95–1.05×
attributes.length	0.97×	0.90–1.05×
item() with varying index	0.99×	0.97–1.01×
Iterate four attributes	1.01×	0.99–1.02×
outerHTML serialization	1.01×	0.97–1.03×
Create element + materialize empty map	1.10×	1.06–1.13×
Create element + set first attribute + read length	1.10×	1.07–1.13×

Most established-operation medians remain close to baseline; getAttribute was noisy and is not claimed as an improvement. Both measured creation paths are about 10% slower, so the change is not performance-neutral across all operations. An additional independent head/base spot-check reproduced that creation regression.

Compared with the current shared fallback, the scratch per-instance proxy took 1.78–8.15× as long for the tested get/set/remove/length/item/iteration workloads, 1.08× for serialization, and 1.37× for creation + first attribute. It was not slower everywhere: indexed property reads were about 9% faster, while named hits were about 11% slower. This supports choosing the shared fallback for existing method paths rather than claiming universal superiority.

Heap

Measured post-GC process.memoryUsage().heapUsed slopes over 0, 25,000, 50,000, and 100,000 retained elements, each with a materialized map. Five fresh processes per variant/shape; five full GCs before each reading; equal preallocated reference arrays and one separately retained Window. These are fitted used-heap bytes per element + map (+ attributes), not exact object retained sizes or RSS; a fixed shared-proxy allocation is not a per-map cost.

Shape	Base	Shared fallback	Per-instance proxy
Empty materialized map	~352 B	~352 B	~384 B
Four attributes	~992 B	~992 B	~1,024 B

No per-map slope increase was measured for the shared fallback versus base. The per-instance proxy adds ~32 B/map in both shapes, reproducing the earlier heap finding on the current code.

Index parser

A separate five-round property-read comparison tested the current parser against three semantically equivalent variants: bounds-first checks, unsigned-bitwise validation, and canonical-string equality before integer/bounds checks. Actual built parser exports passed accepted/rejected input checks, including symbols, fractions, aliases, numeric inputs, and array-index boundaries.

Bounds-first and bitwise variants were within ±5% of the current parser by median across the four workloads, with overlapping/noisy ranges. String-first matched indexed reads, but took 1.04× as long for named hits, 1.06× for missing names, and 1.20× for rejected numeric-looking keys (five-pair range 1.17–1.23× for the latter). This supports rejecting invalid numeric forms before string coercion; it does not establish that the current expression is universally fastest.

Bundle size

Fresh pnpm size --json runs after isolated frozen-lockfile installs measured the repository's polyfill Brotli fixture at 10,893 B on base → 11,060 B on head: +167 B (+1.53%). Both pass the unchanged 12,500 B limit. These replace the old branch's inapplicable size delta.

Web Platform Tests

• Enable six supported cases in dom/nodes/attributes-namednodemap.html.
• Keep two cases deferred because they require Document.createAttribute(), with one also requiring an exposed NamedNodeMap constructor.
• Promote the Document-getElementById.html case that mutates an ID through element.attributes[0].value.

Intentional scope

The shared prototype fallback supports ordinary indexed and named reads, not a complete Web IDL legacy-platform-object implementation. It does not add has, own-key/descriptor reflection, assignment/deletion restrictions, alternate Reflect.get() receiver behavior, or indexed precedence over properties placed directly on a map or on NamedNodeMap.prototype.

Document.createAttribute() and global NamedNodeMap exposure remain outside this PR. Existing namespace normalization, HTML case-folding, mutation invariants, and reactions on main are preserved rather than reimplemented.

Validation

• Focused NamedNodeMap tests: 24 passed.
• Full unit suite: 618 tests passed across 44 files.
• pnpm type-check, pnpm lint, pnpm build, pnpm size, and git diff --check: passed.
• Playwright: 13 tests passed on an unchanged retry; the initial run had four iframe interaction timeouts.
• Classified WPT: 13 files passed, with all 78 supported cases passing and 23 expected deferred failures. Used the cached pinned WPT assets.