Indeed. But I do think your approach - which means that the binary data is actually interpreded as a datum of a specific type, drastically increases the risk.
Agreed.
As I noted earlier, I don't at all think deTOASTing is a must for this function to be
valuable, just as tuple_data_split() is also valuable without deTOASTINGing.
I believe "best effort" is very reasonable in the case of a what is an investigatory