It’s because of content negotiation *
With the .html.heex extension, phoenix will generate a render("foo.html", ...) clause in the view module. When rendering a page in the controller, passing :foo instead of "foo.html" or "foo.json" will make phoenix pick the correct clause based on the request headers.

The .html part of the name is used for content negotiation, and the .heex part is used to tell which compiler to use.

Content negotiation - HTTP | MDN (mozilla.org)
Phoenix.Controller — Phoenix v1.5.12 (hexdocs.pm) (rendering and layouts)
Phoenix.Controller — Phoenix v1.5.12 (hexdocs.pm) (render examples)

* That’s my conclusion, there may be other reasoning behind that

EEx is completely ignorant of the surrounding text. It might as well be binary data I think. HEEx actually parses the HTML tags and will raise an error if the syntax is incorrect. One great advantage is, for example, a more concise interpolation syntax in the attributes which is much more friendly to the eye:
<.form for={@user} id={@form_id} {@add_attrs}></.form>

Alright, let’s say the new Phoenix release had used the html.eex file ending. How would tools like editors, GUI Git tools etc. know what they’re actually dealing with? In an html.eex file attribute interpolations with braces wouldn’t mean anything. But if we have the particular extension heex then the tooling can understand that it’s dealing with something that’s more than EEx.