The model CodeMirror edits is flat text. It can apply styles (including variable fonts) that are a pure function of the text, and there are APIs to imperatively apply extra styling and even insert "widgets" (arbitrary DOM elements) and hooks to react to events. This lets you (ab)use CodeMirror for semi-WYSIWYG editing reasonably easily, but it works best when you strive to keep the rendering a function of the underlying text. E.g. Firepad abuses CodeMirror for WYSIWYG, which generally functions well, but after a couple years they're still debugging some problems with copy-paste. And I think they only pulled it off by maintaining a well-defined WYSIWYG data model underneath CodeMirror (they needed it for syncing).
ProseMirror's model is structured. It's a nesting of elements strictly conforming to a configurable schema (by default close in spirit to Markdown - paragraphs/lists, emphasis, links, etc). Note that it's not "flat WYSIWYG" where each character/line has a style; a sub-list is actually inside the parent list, not just a bullet which happens to have higher indentation.
But, then again, shouldn't a single editor be able to handle those two usage-domains? (As this would mean the editor would be more powerful in terms of extensibility).
Code is line-oriented and usually monospaced. Prose is not, and usually requires local layout (spacing, alignment and so on), embedded media (such as images) and variable typography.
