feat(cloud): reclaim realtime connections via foreground lease

[Neonforge98]

Summary

• Foreground lease for the billable Supabase Realtime socket: focused window holds the connection; hidden releases immediately; blurred-but-visible releases after a 45s one-shot grace timer (cancelled on refocus). Cuts fleet peak-connection billing without the alt-tab teardown/recovery storms of a zero-grace release, and keeps presence viewer chips stable across focus flips.
• Event-driven scheduling only — the recurring 60s/5min cloud pass is gone. Outbound pushes (EventStore writes, project outbox) now run even while hidden, so a minimized agent run stays visible to teammates.
• SUBSCRIBED-edge recovery policy (org2CloudRealtimeRecovery.ts): short disconnects (<5min) and rejoin storms (<30s cooldown) recover via delta cursors; only long gaps run the full listings + entitlement refresh. Full recovery force-fetches the active session's comment thread past the 30s TTL.
• Retry hardening: jittered reconnectAfterMs (fleet was phase-locked on realtime-js's fixed schedule), exponential backoff for the comments error loop (10s → 5min cap).
• Restart write-storm fix: seedFromRemoteSummary now seeds the full-payload metadata hash (and the missing avatarUrl), so restarts no longer re-upsert byte-identical metadata for every pushed session.
• Full-listing scoping: start/online/roster recovery issues the complete collab listing only for the active org; background orgs use their delta cursors and get an authoritative listing on activation.
• UI: dropdown option lists restore their visible-scrollbar styling (scrollbar-width: thin made newer WebKit ignore the ::-webkit-scrollbar rules — an overflowing sidebar org selector looked complete at 4 rows).
• Docs: full three-part scalability audit + fix log under docs/cloud-scalability-audit-2026-07-23/.

Server-side counterparts (lock-order fix, O(1) storage accounting, materialized comment counters, signal debounce, indexes) shipped separately in ORGII-cloud-infra as migration 0003_scalability_p0.sql.

Test plan

• tsc clean; Org2Cloud suite 659/659 incl. new tests for the lease grace state machine (fake timers), recovery decision policy, and seed-suppresses-upsert regression.
• Real-machine verification on the dual-instance setup (API Calls panel + ps sampling): idle CPU 0% / RSS stable on both instances, 8 focus flips with no socket churn, one bounded burst on cloud-org activation then complete silence, 0 likely polling.

🤖 Generated with Claude Code