I do believe there is a stronger argument for using text attributes for spelling/grammatical errors. These do seem to be more similar to the current use of text attributes to me and are definitely presented in a similar way. However, if other ATs do need to access the possible choices using actions, etc., this may not be suitable.
Our argument for using text attributes as much as possible is that this way, we can pick up all the meta information associated with a given run of text in one shot. For a screen reader, we generally want to simply read some or all of this meta information when reading or moving through the text.
Ultimately, as long as the information is accessible, we are happy. However, it'd be good to make that accessibility as intuitive and consistent as possible.