While I like the idea too, I have the same reservations as @morbrorper about it not becoming absolutely impossible. There are some really odd non-literal constructions and translations, and it would require endless searching trying to even get a lot of sentences started, let alone being able to handle the various expressions. Perhaps all the current linked sentences could be alternative answers, but that will lead to you probably always picking the one you’re comfortable/familiar with, as well as indeed not covering as extensive of a range as e.g. Duolingo like @morbrorper also said.
I’m wondering though if the new grammar breakdown feature might come in handy here though if wanting to implement something like this. You could at least be shown the expected construction of [adjective] [noun] [adverb] … to have an idea of what the sentence should become like. It could perhaps have corresponding text boxes for the amount of words expected, and also with the appropriate box widths. This might also allow for preventing accidental submission of incomplete sentences and be able to evaluate it on a word-by-word basis. In this sense it could perhaps become a more useful implementation of the Great Translation Game?
Another thing could be to (only?) have it for the Listening mode (as a toggleable option), to be able to habituate yourself to typing out the full sentence (just like I like to speak out the full sentence in the speaking mode), which might already give you a better overall sentence construction retention.
However, I find that the current system works quite well for me too. You see each sentence multiple times, not just as you “review” the cloze word multiple times, but also with different cloze words from the same sentence, and then with the different alternative cloze options in the same kind of sentence construction. So as you’re getting further along, I’ve really found that you do passively absorb a lot of sentence construction rules, and genders of words. For example, I realised that nowhere do we specifically practice genders (e.g. article + noun), only in the context of the sentences, but I’m still finally somehow getting a handle on the fact that it’s “il sistema”, “il problema”, etc. which would normally be the “tricky” gendered words.
It of course all depends on how you use Clozemaster in the first place, how much you get out of it in the end. I don’t like having the translations enabled by default, preferring to focus on getting the cloze from the context of the source sentence, thus already paying close attention to it. Then I’ll listen to the audio at least once, or a couple of times. Ideally I’ll do the sentence first a couple of times in Listening mode, thus having listened to the complete sentence already a number of times, and then review with Speaking mode, speaking out the whole sentence.
But for example I could imagine there’s a proportion of users too who have translation always visible, and perhaps use multiple choice all along, no audio playback enabled, even after having solved the cloze, and without paying much attention to the whole sentences structure, just treating it as a word by word translation for the cloze. Which is completely fine too, and which is how I’d approach a language when I didn’t know all that much about it yet.
But in my opinion there is already quite a lot of room for picking up lots of things from Clozemaster with the current range of settings, but it does mean you do have to stick with it and work your way through quite a lot of the corpus in order to benefit from it. Clozemaster to me isn’t really a method for instant grammar success, but rather a long beneficial process of subtle exposure.
It only “works” in Duolingo due to the course creators having defined a set of possible translations, and yet it often fails, when people venture outside that set.