Ok, you're seeing the combined interaction of kern/11983 and
kern/11989

I've been testing a fix for 11983 on my laptop for the past month and
a half, and haven't seen any odd interactions.

The particular problem is that, because you have a large number of
very small cylinder groups, some of them are *completely full*, but
have few or no directory inodes.  This causes ffs's directory layout
algorithm to spaz in the way you observed.

the fix is to pay attention to both inode count and free blocks when
deciding where to put a new directory inode.

However, filesystems with geometry like that are just generally bad
news..

						- Bill