I think your best bet on this is the Pmatch question type. It is exceptionally powerful at the expense of being complex. You need the power though to deal with the range of possible answers from students. I don't think the standard short answer question type is going to cut it for you because of the degree of complexity in the language.
But that seems like a rather odd set of additional limitations. In fact, that seems positively overbearing. If I were going to try to do this I would
- Write the Pmatch questions as best I could on a first pass.
- Do a trial run with the questions and actual students (or perhaps teachers posing as students).
- Make sure that all those answers were graded suitably and refine your Pmatch questions to correct any places where they weren't graded well.
- Run the entire quiz with the actual student set. Keep the review options set so that students can't see their results at all.
- Once the quiz has been completed, use the question statistics to review the answers to each question in bulk. Check again to make sure that things are graded well. You're not looking at each individual student's paper this way, only the bulk collection. Regrade and then review the statistics again until you're happy with the accuracy.
- Change the review options so that students can see their grade.
I could also see skipping straight from 1 to 4, but it easier to work with a smaller data set first.
The other thing to keep in mind with things like this is that the upfront investment is large, but once you've done it successfully then you can reuse this thing next year/semester with only minimal changes.