Empty the compliance database with one script

[lemenkov]

prepare_test_schema.py builds the accounts the suite needs. Nothing emptied them afterwards.

The suite compares whole object lists — ComponentReflectionTest asserts on the set of tables a schema contains — so tables left by an earlier run fail tests that have nothing to do with them. CI never meets this because it starts a throwaway database per run; a database you keep accumulates, and the local pass count drifts away from CI's for reasons that look like driver bugs.

Measured against a database used for a day of work, through Mirror-over-PostgreSQL:

before:   50 failed, 963 passed
after:  1013 passed, 0 failed      <- the number CI reports

That was the first time the documented baseline has been reproduced locally.

How it works

Objects are found and dropped through the Oracle data dictionary over an ordinary connection, not with SQL for one particular backend, so the same script resets a real Oracle and a Mirror standing in front of PostgreSQL. Verified on both.

It only reports by default

Not ceremony. An earlier draft queried an unqualified user_tables and offered to drop pg_publication, pg_largeobject, sql_features — 103 objects, the server's own catalogs. The dry-run default is what caught it.

The dictionary lookups are schema-qualified for that reason: against Mirror-over-PostgreSQL, sys.user_tables answers with the 4 real leftovers where the bare name answers with 73 rows of PostgreSQL catalogs. That difference is a Mirror gap and will be reported separately — applications write FROM user_tables, not FROM sys.user_tables.

Two fallbacks for cross-backend behaviour

• CASCADE CONSTRAINTS then plain DROP — Oracle will not drop a table a foreign key references without the former; the Mirror rejects that spelling (ORA-00900).
• Bare name then quoted name — the dictionary reports names in Oracle's folded upper-case form, while the PostgreSQL table is really lower-case, so quoting asks for an object that does not exist. The quoted spelling still follows, for the mixed-case and reserved-word objects the suite creates on purpose.

Honest reporting

The run reports how many of the objects it found were actually removed, and exits non-zero when those differ. An earlier version of this script printed "4 object(s) dropped" when all four had failed — it counted what it found. A reset that claims success it did not achieve is worse than one that fails loudly, because the next suite run inherits the mess and blames itself.