Verify live_write notification teardown by end state

[millsmillsymills]

_delete_notification_any_state decided whether a seeded notification leaked by counting failed deleteNotification mutations. The bin holding the notification is unknown at teardown, so both bins are tried and at least one call is a no-op — and the API may report a no-op either as success or as an error. The count therefore tracks error conventions rather than whether anything leaked: it raises a CLEANUP FAILED banner for a notification the delete test already removed, swallows a real failure when the wrong-bin delete no-ops successfully, and surfaces only failures[0], which is the benign wrong-bin error.

Changes

• delete_notification_any_state reads the end state back after both delete attempts and reports an orphan only when a bin still holds the id. Both bins are paged to the end, since QUERY_NOTIFICATIONS requests no sort order. Every delete error is attached to an ExceptionGroup, so both causes reach stderr through run_cleanup. A read-back that itself fails is reported rather than treated as success.
• The session-end orphan scan read structured_content as a bare list. For a tool returning list[Notification] FastMCP always returns a {"result": [...]} envelope, so the isinstance(notifs, list) guard returned early on every run and the scan had never reported an orphan. It now unwraps the envelope through select_mcptest_orphans, lists ARCHIVE as well as UNREAD, passes an explicit page limit instead of the tool's default of 50, and still names any orphans it collected when a later bin fails to enumerate.
• build_raw_client and the teardown helpers move from tests/live_write/conftest.py to the new fixture-free tests/live_write/_teardown.py, following _gates.py: importing the conftest registers its session-scoped autouse fixtures and would skip the run.
• tests/unit/test_live_write_teardown.py drives the helper offline against a fake client that models both wrong-bin API behaviours. It covers an already-deleted notification not being an orphan, a single real failure alongside a no-op success not being swallowed, both delete errors reaching the report, an orphan detected in the archive bin and past the first page, single-page bins costing one request, an unverifiable end state raising, and the client being closed on the raising path. A separate test pins the structured_content envelope the orphan scan parses, so a FastMCP change fails in CI rather than silently emptying the backstop.

Eight of the nineteen tests fail against the previous implementation.

tests/live_write/ remains gated behind the live_write marker, UNRAID_ALLOW_LIVE_WRITES=1, and the mcptest* name guard, and is deselected in CI.

Closes #296

🤖 Generated with Claude Code