- Re-raise SlidingSyncUnknownPosition from finish_request. A racing
request on the same connection can clear the position before the
persist, and Python would fail the same way, so a Python retry only
repeats the work.
- Treat an empty events stream cache as a miss, as Python's
get_entities_changed does, so Rust asks the database.
- Copy the room statuses out of the connection state before building
Python objects, so the lock is never held across a Python call.
- Trace the rooms Python computes for Rust, as the Python path does.
- Test the per-room fallback: Python computes a room Rust leaves to it,
and its changes reach the Rust state.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01L51DoxB8VnH8UjXJFDgpNG
