Well, what I did was:
- Take out sdg.
- Reformat both caching and backing devices and started over.

I still have the kernel logs of the previous fault condition, but
not the data anymore.

What I notice are:
- The 100% CPU usage for four bcache_writebac processes is back.  I've
  already tried raising the value you suggested from 30 to 60, but I did
  not have an opportunity to reboot yet.  Will report back on this.
- When I run bcache-super-show on the backing devices, it works just fine.
- When I run bcache-super-show on the caching devices (while in service)
  I get something like this:

sb.magic		ok
sb.first_sector		8 [match]
sb.csum			FFFFFFFF5D55BF8D [expected 726D3B42D08769EC]
Corrupt superblock (bad csum)

  On both caching devices.  Is this to be expected?
-- 
Stephen.