[trinity-ability]

Summary

A full tests/run-full.sh run on a local stack failed several tests that pass when run alone. Every one of them is a test-side problem: order dependence (pytest-randomly shuffles the order when it is installed), state that leaks between tests, or an assertion that went stale after an intended product change. This PR fixes them in the tests only. No product code changes.

The two real product bugs from the same run were filed separately as #3676 and #3677.

Integration / API tier

Test	Cause	Fix
test_nevermined_permissions.py::test_get_config_no_config, test_nevermined_payments.py::test_info_no_config, test_shared_folders.py::test_folders_default_values, test_read_only_mode.py::test_get_read_only_status_default_is_disabled	Each asserts "fresh agent" defaults on the module-scoped created_agent, which sibling tests change. Under a shuffled order the change can come first.	Use the per-test isolated_agent fixture.
test_agent_rename.py::test_rename_agent_name_sanitization	It always renamed to test-agent-with-spaces. Deleting the renamed agent soft-deletes its ownership row, and that row keeps reserving the name and its volume base by design (#1664/#1671). Every run after the first got 409.	Derive the target from the fixture agent's unique suffix.
test_subscriptions.py::TestSubscriptionAutoAssign round-robin and tiebreak	Since #2409, auto-assign ranks by cached headroom. A real operator subscription has a reading and the suite's fake tokens never do, so it always wins.	Skip when operator-owned subscriptions exist, the same guard the no-subscription test already uses (#2420). The tiebreak subscriptions now carry the suite's test- prefix.
test_operator_queue.py::test_respond_to_pending_item	Answering a platform-raised item lands acknowledged, not responded (#2372). The test picks whichever pending row comes first.	Pin both shapes. acknowledged must also set acknowledged_at.

Unit tier

Each fix is in the test that leaked state, not in the test that failed.

Victim	Polluter and mechanism	Fix
test_ent814_served_seat.py	test_brain_orb_postprocess.py: dispatch_postprocess lazily imports task_execution_service while database.db is patched. The unit conftest evicts that module after each test, so its fresh import bound the stub.	An autouse fixture pre-imports the service before each patch.
test_2973_depth_new_roots.py	test_monitoring_dormant_skip.py: a bare sys.modules["services.monitoring_alerts"] = stub that was never undone.	monkeypatch.setitem.
test_2572_credentialless_adoption.py (apply-phase restarts)	test_subscription_auto_switch_no_cred_import.py: the restore fixture fixed the sys.modules slot but left a stale module on the parent package attribute. The victim then patched the wrong copy of _restart_agent.	Snapshot and restore the parent-package attribute too.
test_ec_loop_service_edges.py::test_m16_infinite_budget_never_trips[postgres]	Not pollution. On PostgreSQL the cost columns are float4, so writing 1e300 raises NumericValueOutOfRange. That is the known bug #3441.	Strict xfail on the Postgres variant only, citing #3441. It will flip when the bug is fixed.

Test plan

• Each unit polluter/victim pair fails before the fix and passes after, run with -p no:randomly.
• The touched unit files plus their victims pass under -p randomly with seeds 1, 7, 42, 12345, 987654 and 2026: 188 passed, 1 xfailed each time.
• The seven touched API test files pass against a live local stack under -p randomly. Seed 1 gave 149 passed, 3 skipped. Seed 424242 had one transient ReadTimeout on the first run and gave 149 passed, 3 skipped on the re-run.
• The [postgres] xfail was not executed locally, because it needs the disposable Postgres tier.

The 3 skips are the subscription precondition guard. They fire because the test stack has a real operator subscription.

🤖 Generated with Claude Code

[vybe]

merge-train: batch validated on train/20261013-2303 (#3685)