I'm observing a variety of task/test failures that seem to be provoked
by OS crashes.

The problem is most easily observed on quawks, because it manifested
as a surge in provision failures:
    http://shortn/_werCzNE2PZ

The provision failures themselves show histories like this:
    2018-06-20 16:11:29  OK http://cautotest.corp.google.com/tko/retrieve_logs.cgi?job=/results/hosts/chromeos4-row10-rack8-host21/807709-provision/
    2018-06-20 12:59:44  OK http://cautotest.corp.google.com/tko/retrieve_logs.cgi?job=/results/hosts/chromeos4-row10-rack8-host21/806927-repair/
    2018-06-20 12:50:55  -- http://cautotest.corp.google.com/tko/retrieve_logs.cgi?job=/results/hosts/chromeos4-row10-rack8-host21/806884-provision/
    2018-06-20 04:11:05  -- http://cautotest.corp.google.com/tko/retrieve_logs.cgi?job=/results/210164288-chromeos-test/

Key features of the sequence above:
04:11:05  The DUT finished testing, and then went idle until the next suite
          was scheduled.
12:50:55  Initial provision failed, because the DUT had gone offline in the
          interim.
12:59:44  Repair brought the DUT back online (by resetting with servo).
16:11:29  Next provision task passes, and testing resumes.

I _believe_ that this problem is more widespread than quawks.  However,
I don't yet have the data that shows it reliably.