There were two main problems:
1. The unit tests for upgrading adaptive quiz attempts had slighly the
wrong $expectedqa, and so matching that the upgrade was doing the wrong
thing in certain situations. The main issue was that it was setting
-_try = 1 on the first step, which broke the penalty calculation when
the quiz was regraded. There were also some other subtleties with
incrementing -_try that were not right before.
2. It was possible in 2.0 and earlier for two question_states to get the
same seq_number, and restoring 2.0 backups was rashly assuming that that
was unique.
Tim Hunt
0366ef262a