There are a few things I think might make sense, but can leave that for after merging:

• Refactor {HASH reinit and configure-and-reset-context and restrore-context-from-self-if-any} into a single function (used both in finalize and update)
• Check whether we always have to reinit the HASH if we always leave it in a good state
• Consider whether we want to have a debug_assert-only step where we read if the lengths always line up with NBEW (we know it works well for stm32wba55, but sooner or later we'll have chips on the almost-same peripheral, and then that'll be a good debug helper)
• Provide an implementation of update-and-finalize to save the reinit and load-and-restore (but we'd have to add that to the trait in the first place, and due to UPDATE_MULTICHUNK=true, we only save O(1) and not O(n) blocks)

If you agree that none of those has to happen now, I'd merge #23 and open a tracking issue with the above for potential later enhancements.

Sure! I would prefer to merge #23 the way it is right now and add this as future issues, (can be assigned to me)

I agree with all points, just didn't fully get "Check whether we always have to reinit the HASH if we always leave it in a good state". You mean checking the registers before writing on them?