"So: one cell fails -> a whole new batterypack will be needed.
One more reason why the cells will have to be very durable."
For sure - matches up with Elon stressing in ER that battery yield, quality and longevity is really important.
But - are we all that sure that just one or a couple cells failing is a big problem?
I recall various videos from Jack Richard from EVTV where he praised both individual cells and the battery pack as a whole after taking them apart and testing them rigorously.
As I recall Tesla has the least amount of variance between individual batteries he ever tested. Also, the batteries were grouped and various monitoring and error handling was added, both in software and hardware on both the modular and pack level, enabling very good pack performance and longevity despite minor battery variation and/or a few failing batteries, out of the ~3000 cells in the 2017 pack.
(I do think that the 4x modules per pack design was a conservative hedge by Tesla: They could replace one of the 4 subsections instead of a whole pack.)
The structural pack and 4680 is a whole new deal of which we know very little. But, it would make sense if Tesla employed some of the same methods they have developed over the years. It could perhaps be possible to fabricate a top layer (to glue to the cells) of both conducting material and perhaps also with embedded measuring hardware, enabling one level of error handling, as well as serving as a structural member - a kind of very large and rigid circuit board.
Another possibility is that they made the conscious decision to skip one or two layers of 'error handling or mitigation' techniques in order to drive down the cost of structural batteries.
If so they would likely only do this if the error rate on the pack level is tested to be really low.
How to determine this reliably? Here they have the long-standing cooperation with Jeff Dahn from Dalhousie to draw lessons from: One of his great contributions is to measure batteries very precisely, enabling continuous testing to yield useful result with a high degree of reliability in time frames of a few quarters to a year, instead of 5-10 years.
Tesla likely has testing rigs for the new 4680 packs where they can elicit useful quality assessments quickly.
To speculate further, Tesla might very well have a wide range of variants of 4680 batteries and structural pack assembly method. They could then test both the battery yield and longevity and structural stress of the pack of all these variants.
Even though the testing methodologies developed by Dahn et al yield results quickly, the measurements still gain in reliability over time. So it all comes down to how much risk you are willing to take, how well you can pivot and recover from mistakes and sunk cost from earlier decisions that turned out to be not right enough vs the opportunity cost of not acting fast, based on incomplete information.
Luckily, Team Elon has a lot of experience in navigating fluid and complex problems like that.
Edit: Chenkers beat me to it.
What you and @Chenkers are saying makes a lot of sense.
My thinking comes from articles from Gruber that one cell could degrade the performance of the whole pack. But maybe that was about the old roadster.
Apart from making the 4680 cells very durable, I think you are right that we can expect Tesla to handle the problem of cell variance too in tight fabrication tolerances, software and electronics.
The new structural batterypack has a lot of consequences.
But, as we agree on, Tesla team has a lot of very clever engineers to handle those challenges.
And I am sure they will keep on innovating, that really is something in the core of Tesla.