Summary

The Hindsight memory provider can submit async retain work after shutdown begins, causing RuntimeError: cannot schedule new futures after interpreter shutdown during Python teardown.

Root Cause

shutdown() joins existing background threads and closes the client, but there's no lifecycle gate preventing sync_turn() from submitting new asyncio.run_coroutine_threadsafe() calls after shutdown starts. A late sync_turn() call races against the interpreter's cleanup of the event loop machinery.

Fix

1. Lifecycle state gating

• Added _shutting_down, _closed flags and _state_lock
• sync_turn() returns early once _shutting_down is set
• _run_sync() accepts allow_shutdown=True parameter — normal callers are blocked after shutdown starts, but the shutdown path itself can still flush/close

2. Extracted _submit_retain() helper

• Retain submission logic extracted from sync_turn() so both normal sync and shutdown flush use the same code path
• Tracks _last_retained_turn to avoid duplicating a full-session retain during shutdown flush

3. Shutdown final flush

• Before closing the client, shutdown() checks for un-retained pending turns and flushes them via _submit_retain(allow_shutdown=True)
• This ensures recent conversation turns aren't lost on shutdown

Changes

• plugins/memory/hindsight/__init__.py — lifecycle state, _submit_retain(), guarded sync_turn(), enhanced shutdown()

Fixes #15497