Although you lay out the aspects of your question as sequential, I think they're not quite so linear. I will, however, address the different aspects in order.
1. Unless you can be absolutely certain of data protection and that use for transcription is explicitly permitted, I would use for no more than individual words or at most clauses. Optical character recognition is done by pixel, and handwriting is a big challenge; multi-modal LLMs that are trained to handle images struggle, so the chances that a lot of errors will occur is high. A strong language dependence may also be expected, and the 'difficult' cursive handwriting will be particularly liable to LLM errors, so your biases are not overcome. See, e.g., https://arxiv.org/abs/2502.20295 and https://arxiv.org/abs/2503.15195
2: I have to ask: But why do you need the LLM to group responses? My undergraduate lecturers once told me that the way to mark scripts, whether homeworks or exams, quickly is to mark by question. One goes through the stack several times, but one marks much more quickly because one only has to remember the answer to the current question (or even question part). Let me also add that, in spring 2026, a colleague at an institution that is generally pretty 'pro-AI' experimented with using an LLM ('sandboxed') to mark student reports and abandoned the effort after a few hours: it was more frustration and work to get the LLM to provide the feedback he wanted than it was to just sit down and work through the reports as he had for the many years before 'generative AI' products could be claimed to be sufficient for such tasks.
3: Here is where I think LLMs can be helpful, though I am going to start with a warning. LLMs work on the 'norm'; the average, and is what automatically generated feedback will do (or, if marking against a model answer, not recognise equivalent approaches or different solutions). Do you really want your students to be dragged to some mediocre average? Or punished for taking a different approach and showing something equivalent to the required work?
However, I can see that LLMs may help with one thing: creating feedback comments that you only need to edit a little. So you could start your marking knowing (from experience) the things that you tend to want to write something quite 'unfriendly' in response to (i.e., the things that frustrate you), and spend a few minutes developing some templates for you to adjust as necessary.
Another strategy I would suggest here, and irrespective of whether you use LLMs at all, is to flick through the scripts before marking: if they are all getting the same thing wrong (and especially if they do it in the same way), you can develop a common, specific feedback. Depending how long and detailed that feedback is, you might also choose to not provide feedback on that particular question, instead marking it as right or wrong (if wrong, point to the common feedback) and instead provide a single feedback for everyone, pointing out that many students got it wrong in the same way. You will need to be very clear about redirecting students to the 'correct' solution. (If you use the same questions for several years, you could make a copy of a characteristically wrong response and annotate it, saving yourself the work in future years.)
4: You should definitely tell your students if you use LLMs for any part of the marking (or plan to). If you don't you will almost certainly destroy their trust in you (though it's no guarantee you won't destroy their trust in you if you tell them). If you require them to declare LLM usage or advise against it, and then use it yourself, especially without admitting it, you are also being dishonest and hypocritical. Would you have liked your lecturers to have treated your own work with such disdain?