Because case numbers are actually not easy to measure.
I'm basing this of my personal experience working in a German healthcare company, we took up testing during the pandemic; All the tests we do are on unvaccinated people.
That's because unvaccinated people are the only ones who need mandatory, and up to date, tests for pretty much everything from the barber to eating at a restaurant. Recently, it's become so strict that unvaccinated people even need a test to use public transport.
Vaccinated people do not need these mandatory tests, I have no clue who tests them and at what intervals, it certainly ain't the public testing stations, as those are now literally swarmed by the unvaccinated and nobody would go there if they didn't have to.
Which means that in practice we have a massive sampling bias going on; Somebody who's unvaccinated will be very likely to get flagged as positive very quickly, while somebody who's vaccinated would need to decide on their own to do a test.
On tip of that, a lot of states are now on 2G+, recovered or vaccinated and tested (anti-gen), for things like cinemas (requiring official tests in e.g. Bavaria) or self tests on site (fitness studios for example). Kids are regularly tested in school (my son three times, my daughter twice). Sure, some of those tests, those at fitness studios, are not counted. If those are positive, I think the number of people not getting a proper PCR test is rather small.
There is little control theoretic relevance of the Dunkelziffer smoothly varying between 2 and 6 as long as that doesn't make the controller overlook a wave (it doesn't). The sign of the policy should still be the same and if measures don't work, harder measures are imposed means we basically got a PI(D) controller and can achieve stationary solutions even if the Dunkelziffer changes smoothly.
And what causes lack of access? Total number of people hospitalized. And how does that number grow? Exponentially.
So every public health department in the world is optimizing for "How do we prevent our limited number of ICU-capable health facilities from being overwhelmed."
And that comes down to predicting the future hospitalization rate, which can be predicted by the current case count as you said.
And the amount of complex hospitalizations due to COVID can very quickly become a couple orders of magnitude above "normal operations."
Matching with observed reality a better answer may be: 18 months of enhancing the comorbidities of being overweight and in poor physical condition. Sitting on couch while snacking on carbohydrates and fear has been detrimental to many people.
For example, if you implement a travel rule that every traveler needs to get tested a day before they leave and a day after they arrive, then you’re going to see an increase in cases (both real and false positives).
It’s important to be cautious, but it’s also important to not be a know-it-all who just because they can think of a limitation of a data collection / statistical method they can suddenly cast doubt on the whole thing, even though people who know much more about it already probably did exactly the same thing.
Finden a limitation of some data is extremely easy. Extremely easy. It will literally always be possible whenever people are working with data about people because this shit is hard. That doesn’t mean you get to name that limitation, cast doubt on something and piss off. That’s unproductive and irresponsible.
These are the facts: incidence is a great leading indicator if some care is taken with it and it’s regularly put into the larger context (i.e. given more vaccinations a higher incidence is possible until things turn to shit). And you can shout data limitations at that central fact until you are blue in the face but it’s still gonna be there and true.
Why would you want to control a system with 2-3 weeks of additional lag in your input data when case numbers (a thing that is somewhat easy to measure) correlates so well future deaths?
Ideally politician would have some family of models the equivalent of the CDC runs 3 weeks into the future which accounts for seasonality, social events so we could model predictive control on the death numbers but my impression is that is not done (transparently), due to mistrust in experts.