I agree that the status quo seems extremely harsh, especially given that it may not be the validators fault to not have seen the fraud proof (none might have been produced or a network split has occurred). I would say the proposal to increase this to two days would be very welcome, and I also like that it leads to continuity of the beacon chain.
Note that the STARKed data availability root is required for this construction to work. Coded Merkle trees or the original 2D scheme are not acceptable, as in the case where an erasure coded root is fraudulent, different validators may durably disagree regarding whether or not it is valid.
I don’t quite understand this part: Surely this is also a problem in the status quo, where exactly does the new proposal differ so that this is now a fatal problem?