In the end, I only lost two sectors: one in the middle of an ISO file in my home directory, another somewhere in my DocumentRoot. Both files were easily recoverable. The affected file systems are now safely parked on a mirror while I get the new array up and running.
ZFS proved uncooperative at first: I had trouble getting a consistent and up-to-date set of patches, and every time I tried to create a file system and copy data over, it would panic. Pawel and I tracked it down to zfs_reclaim(), and finally figured out that it was caused by zfs_reclaim() calling vdropl() directly instead of vdrop(). The thing is that vdropl() is actually private to vfs_subr.c, and declared static; the code just happened to build and work because ZFS was being built with most warnings turned off, and most testers didn’t set CPUTYPE. Giving vdropl() external linkage and a prototype in vnode.h put an end to the kernel panics.
Oh, and by the way, bunnies are cute.
I guess that led to yesterday’s post. :-)