1. Time in severe hypoglycemia - ideally 0%
2. Time in severe hyperglycemia/diabetic ketoacidosis - ideally 0%
3. Time in euglycemia (also called time in target) - clinical target is >70% and for reference the median healthy non-diabetic is in target ~90-95% of the time.
Closed loop systems are very good at #1 and #2 as it takes a while for levels to get to the severe state and insulin can be administered (or withheld) based on CGM.
When we talk about algorithming out of meal announcements it's whether historical patient-specific blood glucose levels and insulin administrations (i.e. a prediction of what you eat and when) combined with CGM can keep #3 acceptable, not necessarily optimal. Medtronic is using this approach and their newest model more or less eliminates the need for accurate carb-counting but they still require meal announcements. The hope/idea is that this can potentially be eliminated in further iterations.
Another important thing to keep in mind which is sometimes lost in these discussions is that we don't treat numbers we treat patients (i.e. what are the clinical outcomes). Generally speaking, we assume the closer to normal the better but we don't have actual data about how much an extra X% outside of target ranges matters in terms of clinical outcomes and complication rates. We only really started getting this data with CGM and complications in these mild states would require very large cohorts and long (10-20 year) follow-ups to detect differences as they're likely to also be mild.
So while you're absolutely correct regarding the limitations and that an algorithm cannot outperform accurate carb-counting and meal announcements the missing piece is that it may be sufficient. Particularly if said algorithms result in improved time-in-target for patients who aren't good at managing their diabetes and find meal announcements cumbersome.