Revenue records local test isolation incident and verified fixture scope
During additional closure verification, the focused helper selected a loopback URL with schema=revenue_test but used a Prisma client previously regenerated with explicit @@schema("revenue"). Explicit model schema wins over the connection search path. At 2026-09-27 02:11:32 through 02:11:34 UTC, 24 fixture attempts each created a Client, Market, Credit Account and Lesson in shared local platform_local.revenue. Reservation creation failed on the absent closure_revision column. This is 96 unintended local fixture rows, not an unchanged shared mesh. The prior implementation memo described the state before this subsequent run.
No cleanup, reset, shared-schema migration or historical fixture reuse was performed. All 96 IDs and timestamps are preserved in the read-only inventory. These rows are quarantined from rehearsal by process; they are not eligible proof fixtures. Cleanup would be separate scoped work.
Scope evidence
Queries against loopback 127.0.0.1:55434/platform_local found exactly 24 rows each in clients, markets, credit_accounts and lessons in the incident interval, and zero new credit_reservations, orders, order_externals, credit_ledger_entries, external_actions, invoices, invoice_externals, refunds and dispatcher_event rows. These are database observations, not an inference from test failure. The fixture call sequence and 24 test failures match the counts. Provider calls are mocked in this suite and fixture setup failed before provider invocation.
The effective subprocess settings were DISPATCHER_FANOUT_ENABLED=false and EXTERNAL_ACTIONS_RUNNER_ENABLED=false, checked after the test wrapper's environment-file precedence. Both intended database URLs targeted the loopback test schema. The defect was the generated client, not a remote database URL or enabled runner.
Prisma's timestamp-without-time-zone values are UTC wall-clock text. The first JSON inventory used node-postgres date decoding in Hawaii, which incorrectly rendered those values ten hours later. The attached corrected inventory casts timestamps to text and uses explicit timestamp bounds 02:11:29 through 02:11:37. PostgreSQL reports session timezone America/Chicago; a second query covers its corresponding server wall-clock interval, 21:11:29 through 21:11:37 on September 26, and finds zero rows in every surveyed table. This also covers server-default insertion timestamps for the action/outbox assessment.
Containment and preflight
The focused helper now regenerates the test Prisma client before running Vitest. The shared test wrapper rejects any generated model annotation outside revenue_test before loading a test. Positive verification found all 48 generated model annotations set to revenue_test; a negative check supplied the shared revenue annotations to the same guard and confirmed rejection before database access. Both effective DATABASE_URL and DIRECT_URL destinations were independently checked as loopback, port 55434, database platform_local, schema revenue_test.
The corrected isolated run passed 111 tests, including 24 real Postgres tests. This rerun occurred before the coordinator's request to pause further database tests arrived; subsequent database work was read-only inventory. The coordinator then confirmed this is the sole active Revenue task and authorized continued isolated proof after preflight. Cross-domain rehearsal remains paused until the coordinator accepts incident containment and the remaining producer gates. Future Revenue server work must check the generated client state, because generation is global to this checkout.
The runtime fixes and local evidence remain uncommitted in the Revenue main working tree. Unrelated recognition work was preserved. No production data, deployment or runtime flag was changed.