One thing I'd add to this is testing the negative cases for both reads and writes, not just verifying that each user can read their own rows. For a multi-tenant app, I'd test with two authenticated users belonging to different tenants and verify that Tenant A cannot read, update, or delete Tenant B's records. I'd also test the same paths through any RPCs or server-side functions that can access the database. The part I find especially important is making sure the test setup resets the database state between runs. Otherwise, a previous policy change or stale fixture can make the results misleading. I like that you deliberately introduced a permissive policy to check whether the tests actually catch the regression. That's a much stronger test of the test suite than simply confirming the expected queries pass.