On Django 1.7 and higher, we utilize a database backend's

DatabaseSchemaEditor to generate a lot of our SQL. This is done by

putting the editor in SQL-collection mode, which tells it to record the

SQL statements it would use execute rather than executing them.
There's a design issue in SQLite's schema editor in Django 2.0 and

higher. Despite being put in SQL-collection mode, it will try to disable

foreign key checks and then check if it succeeded. If it did not (if,

say, in a transaction), it will immediately fail. This makes sense when

executing SQL, but not when collecting.
No other database backends have a check like this, so the failure case

was limited to SQLite.
There's no way to get around this through the standard usage of the

schema editor as a context manager, so instead we have to wrap it, put

it in the correct state, and collect the results ourselves.
This enables us to get the data we need without these side effects.
All call sites, including those that didn't explicitly put it into

SQL-collection mode (they weren't triggering calls that would have

executed SQL), now use the new wrapper function.

Unit tests pass for all versions of Django.