Regarding storing directory information in containerfiles and only
reading them into memory to operate (and storing MLEs in RVM when you
change them and haven't synced yet):

this is probably a lot of work

it would avoid the current arbitrary length limit

it would move more data to on-disk rather than RVM, and greatly
increase the amount of data one could store in a Coda server.  It
would be slow if it were all in use, but that's not the big problem
with coda (the big problems are max size, security, and occasional
crashes, IMHO!).
 

So I'm thinking of a representation where the directory is primarily
stored on-disk in a container file, but there is a changelog in RVM
keyed by the FID of the directory.  The semantic content is the
on-disk version modifed by the changelog, and when the changelog gets
big it can be checkpointed by writing a new containerfile, and then
truncating the changelog in RVM and swinging over to the new one.