Be careful with what you mean when you say "essentially text replacement" and "same as a literal space".

In HTML, ampersand entities can only be used outside of tags or inside attribute values, for example: <tag key="&value;">&abc;</tag>. Ampersand entities cannot be used as a substitute for tags; for example, &lt;tag&gt; is literal renderable text and not a tag.

This distinction is important because of odd languages like Java, where \uXXXX processing happens before tokenization. So this is a legal program: class Foo \u007B }. (U+7B is left brace.) One consequence is that to put a literal quotation mark inside a string, you can't use \u0022, but you must use \".

Yes, this is why I added the "essentially". HTML does it after tokenization, but before whitespace processing (which implies before a lot of other things that by necessity happen between tokenization and whitespace processing, these steps are nowhere close to each other). But the author seems to vaguely believe/suggest to either process entities at a much later stage, or carry down some "shadow effect" in addition to expanding the entity to its text content. Both of these are complete non-starters IMHO.

(btw, you/anyone have other examples for that Java-style ordering of escapes vs. tokenization? I can think of C preprocessor trigraphs [which are deprecated since C23] but nothing else…)

C trigraphs is a great example that I forgot, because no one uses it and everyone hates it; glad to hear that C23 is finally removing that wart. I'm not aware of other languages with Java-style escapes before tokenization. I only learned the Java example due to the book Java Puzzlers; it has little impact on real-world code because it's considered obfuscation. I don't know the nitty-gritty details of other programming languages that I come in contact with; I just know Java especially well.