✨ feat(rewrite): DOM-less streaming HTML rewriter (lol-html style)

[gaborbernat]

Editing a few nodes in a large document meant parsing the whole thing into a tree first, which costs memory proportional to the document and is wasteful when the task is a forward transform: retag links, lazy-load images, strip trackers, redact text. This adds turbohtml.rewrite.rewrite, a single-pass rewriter that transforms markup without building a tree, the model Cloudflare's lol-html uses to rewrite responses at the edge. Closes #544.

The rewriter streams the existing WHATWG tokenizer over the input and keeps only the stack of currently-open elements. Each open element is a lightweight node whose sole live link is its parent, so the native CSS selector engine that powers select() matches against that spine directly, and a matched element (or a text run, comment, or the doctype) is handed to a handler that edits it in place through an Element handle: set or remove an attribute, insert markup around or inside it, replace its inner content, unwrap it, or drop it. 🚀 Working memory stays proportional to the open-element depth rather than the document size, so a page far larger than memory rewrites in a fixed footprint, and an untouched construct is copied through verbatim, so a rewrite that edits nothing returns its input unchanged.

The one constraint the design imposes is the selector subset. A no-lookahead stream has seen an element's ancestors but not its following siblings or descendants, so only selectors decidable from an element and its ancestors match: type, universal, id, class, and attribute selectors, the descendant and child combinators, :root, and :is()/:where()/:not() over that subset. A sibling combinator, a positional or structural pseudo-class (:nth-child, :only-child, :empty), or :has() is rejected at compile time with SelectorSyntaxError, matching lol-html's own restriction. The explanation/streaming note covers why, and the migration/lol-html guide maps the HTMLRewriter and Element API across.

The whole engine lives in the C extension (_c/tokenizer/rewrite.c) reusing the tokenizer and selector engine; the Python layer is a thin typed shim. A depth guard bounds the open-element stack so a pathologically deep or unclosed input cannot exhaust memory or the C stack the matcher walks.

[codspeed-hq]

Merging this PR will not alter performance

✅ 82 untouched benchmarks
🆕 1 new benchmark
⏩ 18 skipped benchmarks¹

Performance Changes

	Benchmark	BASE	HEAD	Efficiency
🆕	test_feature[rewrite]	N/A	7.8 ms	N/A

---

_{Comparing gaborbernat:feat/544-streaming-rewriter (61379f9) with main (60cca4c)}

Footnotes

1. 18 benchmarks were skipped, so the baseline results were used instead. If they were deleted from the codebase, click here and archive them to remove them from the performance reports. ↩