In the past, I've either dropped into just straight-up HTML for more complicated bits, or done two-pass rendering. (E.g., have input → tool A → documentation generator → docs. We do this to render a few tables in our docs from YAML data files that just contain the actual data in the table. Our doc system includes an "include this file" directive, so the result can be pasted into other prose still written normally.)

It's not completely optimal. I do wish M↓ supported things like admonitions (callouts for warnings, notes, todos) better, in particular, since our docs are rife with those.