All unknown (and, formerly, silently ignored) keywords were removed from
existing tests. This allows us to detect typos.

The steps get split in two, with the authoritative parts getting copied.
When it's possible to INCLUDE, this should get refactored.

MATCH specification was totally ignoring the response so the test did
effectivelly nothing. Fixing MATCH clauses uncovered that delegations
were incorrect and that queries for localhost addresses have to be
allowed for the test to work.