[michaelthyregod]

The quirks mode checks in PresentationalHints.GetDeclarations (once per element) and in CssBuilder.CreateDeclarationWith read IDocument.CompatMode. In AngleSharp that getter maps the quirks mode to its string through reflection on every read (QuirksMode.GetCompatiblity(): GetField + GetCustomAttribute), which costs more than the rest of an element's cascade. In a profile of GetInnerText it was the largest single item.

This PR reads the mode directly where AngleSharp exposes it:

public static Boolean IsQuirksMode(IDocument? document) => document is IConstructableDocument constructable
    ? constructable.QuirksMode == QuirksMode.On
    : document?.CompatMode == "BackCompat";

• Same result: only QuirksMode.On maps to "BackCompat"; Limited (limited quirks) maps to "CSS1Compat", as before. Other IDocument implementations still go through CompatMode.
• I checked whether the string comparison was there on purpose (a workaround): both checks were added together in the presentational hints work, and CompatMode reads the same field, so I found no reason. If you'd rather stay on the public IDocument API, the alternative is caching the result per document - happy to switch.
• New test LimitedQuirksModeHasNoneOfThem: a limited quirks document (XHTML 1.0 Transitional doctype) gets neither the quirks of the presentational hints nor the unit-less lengths of the parser. It fails when Limited is treated as quirks.

Benchmarks against devel (BenchmarkDotNet, 6 alternating rounds, pinned to the performance cores):

	Change	Allocations
DeepTree InnerText 100 / 500	−26.1% / −23.9%	−5.6% / −5.8%
GetInnerText, real-world / selector-heavy	−2.6% / −2.0%	−0.5% / −0.8%
CssCascade RenderTree / ComputedStyle	−2.0% / −1.1%	−0.4% / −0.2%
Render(), selector-heavy document	−5.0% / −2.5% (two runs)	−0.7%
Render(), real-world document	+10.8% / +6.3% (two runs)	−0.5%

The Render() real-world number is a GC artifact, not extra work: the slow rounds are exactly those in which the candidate ran a gen2 collection (one every 8 renders); the baseline had none. The output and the element count are the same, and allocations are slightly lower. A probe without BenchmarkDotNet's forced collections between iterations measured −0.5% / −2.8% for it.

🤖 Generated with Claude Code

[FlorianRappl]

LGTM!