feat: Redis Streams event bus for WebSocket delivery (RELIABILITY-003)

[vybe]

Summary

Replaces the in-process ConnectionManager.broadcast() + except: pass pattern with a Redis Streams transport so short WebSocket disconnects no longer drop events. Keystone for Tier 2.5 simplification per docs/planning/ORCHESTRATION_RELIABILITY_2026-04.md — #428/#429/#307 reuse the same primitive; #408 dissolves once agent-push completion rides on top of it.

• Publisher: EventBus — fire-and-forget XADD with MAXLEN ~10000 (env-tunable via REDIS_STREAM_MAXLEN), bounded outbound queue so Redis latency never blocks broadcast call sites.
• Consumer: StreamDispatcher — single XREAD BLOCK per backend process, in-memory fan-out, per-client asyncio.Queue(256) with drop-and-resync on slow consumers, 3-failure eviction, supervised reader with exponential backoff, 2s graceful drain on SIGTERM.
• Reconnect replay: clients send ?last-event-id=<stream_id>; dispatcher runs XRANGE catchup capped at the reader's position at register time (prevents catchup/live overlap). Trimmed cursors or gaps >5000 → {type: "resync_required"} → frontend full-refetches authoritative state via REST.
• Frontend: both utils/websocket.js and stores/network.js capture _eid from incoming messages and append &last-event-id= on reconnect.
• Public API preserved: 33 existing manager.broadcast(...) / filtered_manager.broadcast_filtered(...) call sites are untouched — the managers are now thin shims over the bus.

Scope discipline

This PR is WebSocket delivery only. Deferred per plan:

• refactor: consolidate ExecutionQueue + SlotService + BacklogService into CapacityManager #428 CAPACITY-CONSOLIDATE — agent-push completion + queue unification (gated on 2-week soak)
• refactor: collapse 9-path cleanup pyramid once agent is authoritative #429 CLEANUP-COLLAPSE — retires the 9-path reconciliation pyramid
• feat: Agent heartbeat push for fast failure detection (RELIABILITY-004) #307 RELIABILITY-004 — heartbeat push (reuses this stream)
• bug: long-running backend→agent task calls dropped when uvicorn worker recycles #408 dissolves once refactor: consolidate ExecutionQueue + SlotService + BacklogService into CapacityManager #428 ships

Changes

• src/backend/services/event_bus.py — new module (EventBus + StreamDispatcher + helpers)
• src/backend/main.py — manager shims, lifespan start/stop, /ws + /ws/events accept last-event-id
• src/frontend/src/utils/websocket.js — _eid capture, reconnect URL, resync_required handler
• src/frontend/src/stores/network.js — same wiring for the collaboration dashboard WS
• tests/test_event_bus.py — 23 unit tests
• docs/memory/feature-flows/websocket-event-bus.md — new flow doc + index row + xref from activity-stream.md

Test plan

• Unit: pytest tests/test_event_bus.py -v → 23/23 passing (envelope shape, id validation, scope visibility, Redis-unavailable fallback, 3-failure eviction, slow-consumer resync marker, monotonic cursor guard, invalid-cursor resync, agent_name inference, update_accessible_agents mutation)
• Live smoke against localhost stack (playwright):
  • Publish → browser receives payload with injected _eid
  • Close socket, publish 3 events offline, reconnect with last-event-id → all 3 replayed in order
  • Malformed last-event-id → connection accepted gracefully, bad id ignored
  • Dashboard network.js reconnect URL includes captured last-event-id
• End-to-end integration test (Redis + backend + client) deferred — the 2-week soak in the plan is the real signal
• Manual: dashboard loads with 10 agents visible, isConnected: true, no console errors from the new path

Follow-ups (non-blocking, captured in /review informational findings)

• I1: Redis-down silently pauses live updates — signal stream_offline to clients (low priority)
• I3: Frontend refetch errors swallowed silently — add console.warn in catch blocks

Closes #306

🤖 Generated with Claude Code