perf(cloud): batched roster entitlements + paged session listings

[Neonforge98]

Summary

Client side of ORGII-cloud-infra migration 0004_roster_and_listing_scalability.sql (one paste, all offline-validatable scalability fixes). Every path is fully backward compatible with pre-0004 backends.

Part 1 — batched org entitlements (audit B1). Eliminates the get_entitlement_state × N fan-out on every roster read (measured live: 7 extra RPCs per org activation and per coarse-signal refresh). Roster rows carrying an entitlement payload seed the coordinator directly (seedOrgEntitlement: floor mirror + TTL stamp); only unresolved rows fall back to the per-org RPC. Malformed payloads degrade via .catch without failing the roster parse.

Part 2 — keyset-paged full session listings (audit C2). Full cloud_list_org_sessions pulls walk p_limit=200 keyset pages until nextCursor disappears, so a large org can never push one unbounded aggregate through the managed statement timeout. Delta pulls stay single-shot. A pre-0004-signature backend answers the first paged call with PGRST202; the client falls back to the legacy call and remembers per endpoint. Runaway guard at 50 pages.

Part 3 — keyset-paged full collab-state listings (audit H3). Same treatment for cloud_list_org_collab_state: full listings walk the server's unified (updated_at, kind, id) cursor (one cursor pages projects and work items together), delta pulls stay single-shot, PGRST202 falls back with per-endpoint memory, 50-page guard. Ascending keyset order makes mid-walk writes safe — an update moves its row toward the unread tail, never behind the cursor.

Server-only parts of 0004 with no client change needed: work-item project-lock narrowing (M1 — steady-state edits stop locking the project row), comments p_since delta (M4 — client adoption deferred; threads are small), service-role GC for collab tombstones and bookkeeping tables (H3/L2).

Test plan

• tsc clean; full suite green (6156 tests / 684 files); 15 new Org2Cloud tests across entitlement seeding, session paging, and collab paging (multi-page walk + cursor threading, PGRST202 fallback + endpoint memory, delta single-shot body, malformed-cursor degrade, page-cap guard, non-signature error propagation).
• Server counterpart validated twice on disposable Postgres 16 (author agent + independent re-run): full 0001→0004 sequence, idempotent re-run, legacy calls byte-identical vs a 0003-only DB (collab state, comments, 6-step work-item upsert battery), paged unions equal unbounded listings, comments delta exact-set, GC horizon/live-item guards, pg_locks-level proof that steady-state work-item edits no longer wait on the project row.

🤖 Generated with Claude Code