Well, I looked at the core dump. The contents of the dinode appears
to be mostly garbage.... values that should not be possible even if it
were in fact improperly reusing an inode.
I suspect that the remaining 'dup alloc' reports are due to the ATA
driver... either buggy hardware or a missing workaround in our ATA
driver for the buggy hardware. I've been running diff -u on a 6GB data
set and rm -rf + cvs checkout tests for over two days straight on both
of my test boxes without a problem.
One thing that might be worth doing ... try changing the DMA mode for
the ATA hard disk using atacontrol to something slower, reboot into
single user, do a manual fsck, then see if the system is more stable.
-Matt