If I have a recipe with correct checksums in, fetch the tarballs, and then change the checksums (either deliberately or accidentally) subsequent rebuilds of that recipe won't re-validate the tarballs. This can result in recipes being submitted that don't actually build. I think that do_extract() should also verify checksums, ideally with a fast-path for the common case where the metadata doesn't change between do_fetch and do_extract.
Its fetcher design that checksums get checked after the file is downloaded. If the file is already present with the "done" stamp, its assumed the checksums were checked and are right. Checksumming the files at do_unpack adds significant overhead unfortunately.
Thinking a bit further there is a half way house on this. We should write the checksum into the .done file and where we already have an existing file, we can just verify the checksum in the file matches. This would prevent the common case of errors.
A post to yocto@ spoke about a related problem: disk corruption. I'm not convinced we should cater for failing disks as the problems won't be restricted to the source cache, but everything that gets written to disk.
*** Bug 6923 has been marked as a duplicate of this bug. ***
This has been fixed in http://git.yoctoproject.org/cgit/cgit.cgi/poky/commit/?id=36e47a99ae6668cb9a601f137f56011b7de13512.