test: harden weak audit assertions (F7-04, F7-06, F7-11, F7-10, F7-05, F3-05)

[lesnik512]

What

PR 2 of the tail program. Strengthens tests the 2026-06-14 pass-3 audit flagged as passing-too-easily (they'd survive real regressions).

• F7-04 — new test_fetch_skips_rows_locked_by_another_transaction: locks 20 of 30 rows in an open transaction, then asserts a concurrent client.fetch promptly claims the disjoint 10. A regression from FOR UPDATE SKIP LOCKED to plain FOR UPDATE would block on the locked rows → the bounded wait_for fails via timeout. (The existing no-double-claim test guards its named invariant but would pass without SKIP LOCKED.)
• F7-06 — assert dispatch_one threads row.id + row.acquired_token (the load-bearing lease guard) into delete_with_lease, not just the connection.
• F7-11 — capture and assert the relayed body on the successful at-least-once retry (a retry relaying an empty/wrong payload would have passed the call-count + row-deletion checks).
• F7-10 — identity-check the FastAPI-injected producer/client against the swapped fakes (the Context(...) annotation string is the load-bearing artifact; isinstance would pass on a wrong instance). Adds TestOutboxBroker.fake_producer, symmetric with fake_client.
• F7-05 — assert the DLQ headers audit column round-trips (off-Postgres guard on the fake's DLQ copy).
• F3-05 — new test_dlq_cte_quotes_adversarial_table_names pins identifier quoting at the one raw-SQL site (table names with embedded " and ;); a refactor dropping format_table would fail it.

Verification

• just test → 509 passed, 100% coverage
• just lint clean (ruff + ty)

Deferred to a follow-up

• F7-07 (NOTIFY-wakeup determinism — fiddly spy on _wait_for_notify_or_timeout)
• F7-09 (telemetry exact bounds — spans 4 metrics/middleware test files)
• F1-08 (route sync dispatch through fake_client.fetch — touches the sync-dispatch path, higher risk; deserves its own focused change)

🤖 Generated with Claude Code