openai/gpt-chat-latest → B (3:1)

Side B fixes a correctness issue in feed catch-up by anchoring implicit pagination to durable ingest order instead of timestamps, preserving behavior across same-millisecond posts and clock rollback, while also enforcing permission-aware visibility and exposing room metadata. It refactors duplicated feed-building into a shared helper and adds comprehensive integration tests for multi-user, private-room, and permission-revocation scenarios, whereas Side A mainly improves pair-selection heuristics and ordering for ranking refinement with supporting tests.