Emacs

My first suggestion would be to use plz for HTTP. Then I'd use cl-loop and pcase to simplify the rest of the code. Here's a partial rewrite with a TODO for further exercise. :)

(defun wikipedia-article-references (subject)
  (let* ((url (format "https://en.wikipedia.org/wiki/%s" (url-hexify-string subject)))
         (dom (plz 'get url :as #'libxml-parse-html-region)))
    (cl-loop for cite-tag in (dom-by-tag dom 'cite)
             for cite-class = (dom-attr cite-tag 'class)
             collect (pcase cite-class
                       ((rx "journal")
                        (let ((a-tag (dom-search cite-tag
                                                 (lambda (tag)
                                                   (string-prefix-p "https://doi.org" (dom-attr tag 'href))))))
                          (cons (concat "doi:" (dom-text a-tag))
                                ;; TODO: Use `string-match' with `rx' and `match-string' here.
                                (let* ((cite-texts (dom-texts cite-tag))
                                       (title-beg (1+ (string-search "\"" cite-texts)))
                                       (title-end (string-search "\"" cite-texts (1+ title-beg))))
                                  (substring cite-texts title-beg title-end)))))
                       ((rx "book")
                        (let ((a-tag (dom-search cite-tag
                                                 (lambda (tag)
                                                   (string-prefix-p "/wiki/Special:BookSources" (dom-attr tag 'href))))))
                          (cons (concat "isbn:" (dom-text (dom-child-by-tag a-tag 'bdi)))
                                (dom-text (dom-child-by-tag cite-tag 'i)))))
                       (_ (let ((a-tag (assoc 'a cite-tag)))
                            (cons (dom-attr a-tag 'href) (dom-text a-tag))))))))

Regarding this:

And yes, I know that I could probably use a library like s, dash, seq, or cl, but I try to keep my elisp functions free of those kind of things

First of all, cl and seq are built-in to Emacs and are used in core Emacs code. There's no reason not to use them. Second, dash and s are on ELPA and are widely used; it's largely a matter of style, but they are solid libraries, so again, no reason not to use them. They don't have cooties. ;)

You don't have anything to guard against a bad response from the server. e.g.

(unless (equal url-http-response-status 200)
  (error "Server responded with status: %S" url-http-response-status))

To position point at the end of the headers:

(goto-char url-http-end-of-headers)

This:

(setq result (cons (cons ...) result))

Is more clearly expressed as:

(push (cons ...) result)

Better yet, you could map over the elements you're interested in and accumulate the results via mapcar or cl-loop. That would obviate the need for the "results" variable.

You could probably shorten things by using the dom-elements function to directly search for the href's you're interested in in combination with dom-parent to get at the parent elements.

Overall your function gets a 65 out of 130 ERU (elisp rating units).

having

          ))
  )))))

is not very lispy