Thanks for your input, Olichant. I think you've nailed it when you say that the original idea was that the "real" clues should make sense after correction, just as correcting accidental misprints in a piece of text should to something coherent. Therefore I'd say that if the surface of either version of the clue can take a hit, it's the misprinted version, though the less sense the clue makes the easier it is to spot the misprint and solve it.
I've found that the biggest problem occurs when the only possible options for placing the misprints alter the part of speech of a word. Noun/adjective changes are the worst!
The Listener notes state that removing extra letters from clues needn't leave something that makes sense, and from what I recall this applies to extra words and extra letters generated by wordplay too. That makes misprints the hardest way for the setter to generate a message. I'm aware that some solvers are tired of misprints and don't want to see them too often - I'm more than happy to oblige!