Gene deletions are expected to some extent as a result of natural mutations.
The author is hypothesising that there should already exist natural virus particles ("in the wild") with key genes deleted that render it harmless to humans, and that these harmless strains may also be spreading in the population. We could possibly identify harmless virus particles with wide-ranging testing of asymptomatic people, and statistically identify what genes (or what parts of the genome) are important for disease.
The author's concept is that this would have potential as a vaccine. It's similar in concept to the polio vaccine, or giving cowpox to prevent smallpox infection.
In my opinion, the author is making several hypothetical leaps of logic that don't make this feel like a workable plan.
The author proposes to test 50,000 people and use this to identify safer strains: in my opinion that wildly underestimates what would be needed for this to have a chance of working.
Since the virus looks to be harmless in the vast majority of people it infects, it will be difficult and dangerous to distinguish strains that are genuinely harmless.
Third no. I am not arguing this at all.
Fourth is pure opinion.
Fifth the same, but as a counter argument sequencing 10,000 strains has found 69 candidate attenuation strains and that was without even trying.
Sixth no. It is actually quite easy to distinguish a harmless strain from a normal pathogenic strain as we can use the high morbidity rate as a proxy for danger.
Saying they are candidate attenuation strains is like saying sugar is a candidate treatment.
[Edited after reconsidering, because it's worse than that:] Saying they are candidate attenuation strains is like saying drinking bleach is a candidate treatment. They are highly likely to still be extremely harmful.
If, hypothetically, you have identified a strain that has had no severe outcomes even in a large cohort of people with known risk factors, then that strain is most likely harmless.
Of course we don't have this kind of data (yet), but the principle is valid.
Right now, we're not looking, but we should be. That's the point.
How could it possible be harmful to check the clinical history of the patient that was infected with these deletion strains?
Gene deletions are actually very common for virus. At their reproductive cycle, the genes have to be kind of "collected" into their shells; a deletion means only that one of the genes wasn't there at the right time.
In the case of the link that OP posted (http://cov-glue.cvr.gla.ac.uk/#/deletion), only non-frameshifting deletions in coding regions are listed. Most deletions listed are one- or two- codon deletions observed in multiple isolates, which is good evidence. Aside from the terminology issues, OP's claim that some of these deletions may result in an attenuated virus is plausible, subject to further analysis of the structural effect of these mutations (which can be modeled in silico). I don't think it would be useful to "go and visit the people who were infected with each of the different deletion strains and investigate what was their clinical outcome" as OP suggests, since that would be confounded by many factors.
Trying to isolate live attenuated virus from the human population in this way carries a lot of risks and potential ethical issues, and amounts to a very low-throughput, difficult screen. The far higher throughput, better researched method that labs are prepared to do is to infect an animal (possibly genetically engineered to be more human-like in its molecular machinery) and wait for the virus to evolve to target that animal, which will make it less virulent in human. However, for SARS-CoV-2 this is especially tricky because of how explosively fast its replication cycle is. Most vaccine trials for coronavirus are working on presenting the human immune system with pieces of the virus instead of live attenuated virus.
I would encourage everyone in this thread to read the scientific literature on this, which is abundant and freely available, for example this review: https://www.nature.com/articles/nrd.2015.37.pdf
Just a small correction on the 69 deletion strains - they are all in gene coding reasons, but they are both frameshift and non-frameshift mutants.