Problems with bzr.dev tonight

[Martin Pool]

Our plans now are to
 - see if this recurs when pulling from the packs.packs branch into a
new repository branched from bzr.dev
 - check the logic for when revisions are reannotated or not
 - write some check-like object to identify which texts in a knit are
not reconstructable
 - and extend that to be able to amputate those versions, leaving the rest
 - ...

Talking this over Ian and I came up with a hypothesis: this happened
when the packs format changed from being annotated to unannotated,
without a synchronized bump of its version string.  (If that did not
occur we need a new hypothesis. ;-)   People might then have been
using an older version of the pack code, which thought they were still
annotated, to pull from Robert's repository in which the annotations
were removed.  Those changes would then incorrectly be inserted into
their knit repo without being reannotated.  If this is true then we'd
expect this problem is no longer reproducible using the current pack
code, and it should teach us to bump versions even for unreleased
dogfood formats.