Image silent data corruption via bitrot can be frustrating. Without a regime to prevent it, it spreads into all backups. And typical workflows depend on backups for eventual migration to new storage which allows any corruption to be replicated via the backup strategy. You end up backing up the corruption, unwittingly.
Which was one of the big problems with the original G4 Xserve. They added it later in the G5 model, but I suspect much of the damage had been done. Most technical people, just ignored it after discovering that it didn't have ECC and the disks were low end ATA rather than SCSI. With the cluster/HPC version the stories changed to how unreliable the machines were.
At that point it seems only the diehards were willing to trust their data to apple server hardware and it died a couple years later.
How do you think I could discriminate between file corruption and the user making an edit to a file?
(Perhaps if the mtime and filesize doesn't change, but the SHA does?)
The location for metadata depends on the application. For DNG workflows, the metadata is a separate location in the DNG file, with separate metadata and data checksums. For other workflows, the metadata is in a sidecar (a per image file), or stored in a database managed by the application.