The purpose of this is data safety. If there ever was a broken model
that corrupted data previously it would have ruined everything if it
coincided with a schema migration. Now this is much less likely.
As I noted in commit 85b9c2459c I noticed
that the testing of the invalidity of invalid struct "deserialization"
was not sound. That was because if one error was not actually thrown,
the test would still be satisfied by the following error. However, if
only that specific thing was wrong, no error would be thrown and an
invalid struct would be accepted.
Therefore I now changed those tests to also check the exception type and
its message. This means that if an exception is missing, it is
immediately picked up.
I also found out that one exception was actually wrong. I had mistyped
`TypeError` as `TypError`. Funnily enough that still causes an exception
to be thrown at that location, but a `ReferenceError` instead of the
`TypeError`. Such bugs should now be more easily noticable.
