Random idea that might improve load performance: initially alienate the entire document, then progressively convert and dealienate it. This should bring the editor into view quickly and allow the first paragraph to be edited quickly while the rest of the page is still processing.
Possible pitfall: reference numbering and other crazy stuff in MWReferenceNode
See T76544: Explore performance gains from progressive (JIT?) de-alienation of content types in VisualEditor for the same idea, but for content blocks.