Improve parser correctness and performance

[ECorreia45]

Summary

Replace the shared-regex parser with a reentrant single-pass scanner, improve the lightweight DOM implementation, add regression coverage for reproduced parser failures, and include a reproducible before/after benchmark.

Why

The previous parser passed its existing tests but failed several important cases:

• trailing text after an unclosed element was appended to the root
• mismatched closing tags corrupted nesting
• > inside quoted attributes broke tokenization
• markup-like content inside style became elements
• namespace caching leaked across parents
• nested parse() calls from callbacks corrupted the outer parse

Module-level regex state, dynamic regex creation, remaining-document slices for scripts, namespace caches keyed only by tag name, and repeated Set-to-array conversion also added avoidable overhead.

What changed

• add a quote-aware, single-pass parser scanner
• keep parser state local so parsing is reentrant
• derive namespaces from the active parent
• handle raw text and matching ancestor closures
• use arrays in the lightweight DOM hot path
• reuse the self-closing-tag regex
• add six regression tests
• add benchmark.compare.mjs for paired baseline/optimized measurements

Benchmark

Node 24, seven paired median samples, alternating before/after order with explicit GC between samples.

Workload	Tokenizer	Default Doc
Small HTML	+31.7%	+8.2%
Attributes	+18.3%	+4.5%
Medium HTML	+36.9%	+0.8%
Large HTML	+32.8%	+12.4%
Script-heavy	+161.4%	+18.5%
SVG-heavy	+3.9%	+4.3%
Malformed HTML	+61.5%	+5.9%
Geometric mean	+43.3%	+7.7%

The bundled parser increased from 11,212 bytes to 13,715 bytes (+22.3%). This is the main tradeoff to review.

Validation

• 42/42 Jest tests pass
• ESLint passes
• Prettier passes
• TypeScript --noEmit passes
• git diff --check passes