After a private communication by Jeffrey Altman and Marc Dionne,
suggesting that the observed problem was not an OpenAFS problem, but a
problem of our script used to simplify (!?) the restore volumes from the
backup, I checked the situation and indeed they were right:
It was just a stupid bug
The backup of our AFS cell is performed using the standard OpenAFS
backup command and we know that it works nicely if the maximum size of
volumes does not exceeds 450-500 GB.
If at least one volumes exceeds that size, the restore of any volume is
unpredictable, for some it works, for others