i suspect it's because the storage i'm using as a staging area has flashdrives between the RAM and spinnning disks, so upon close, the file buffers may get flushed out of RAM to the SSDs but not immediately be available from the SCSI drives.  Or maybe the (inode) metadata table hasn't finished updating yet.  if so, it's not the fact that the cluster is heavily utilized, but the filesystem is.  this disk is expressly for staging cluster jobs.  i'll see if adding a short sleep and retry once upon error solves this problem... but i won't know immediately as the problem is intermittent.  that's the problem with fancy toys; they often come with fancy problems!