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

[vybe]

Summary

The current WebSocket broadcast in main.py uses in-process iteration with bare except: pass error handling. At scale (30+ agents, 50+ WebSocket connections), clients silently miss critical events — collaboration updates, health alerts, execution status changes — with no way to recover missed messages on reconnect.

Solution

Replace the in-process broadcast with Redis Streams as the event bus. Redis is already running. Streams provide:

• Persistent event log with configurable retention (e.g., last 1000 events or 1 hour)
• Replay on reconnect via XRANGE — clients that disconnect and reconnect get missed events
• Consumer groups for at-least-once delivery semantics
• Built-in backpressure via BLOCK parameter on XREAD

Scope

• Replace ConnectionManager.broadcast() with Redis Stream publish (XADD)
• Each WebSocket connection reads from Redis Stream (XREAD BLOCK)
• Track per-client last-seen event ID; on reconnect, replay from that point
• Add stream trimming (MAXLEN ~10000 or time-based) to prevent unbounded growth
• Replace FilteredWebSocketManager broadcast with stream consumer that filters by user permissions
• Remove except: pass pattern — errors now tracked per-client with eviction after N failures

Acceptance Criteria

• All real-time events (agent_activity, agent_collaboration, process events, health) flow through Redis Streams
• WebSocket clients that disconnect and reconnect receive missed events (up to retention window)
• Failed sends to a client are logged and client is evicted after 3 consecutive failures
• Stream auto-trims to prevent unbounded Redis memory growth
• Existing WebSocket event format unchanged (backward compatible for frontend)
• ~100 lines to replace ConnectionManager, zero new infrastructure

Key Files

• src/backend/main.py — replace ConnectionManager and FilteredWebSocketManager
• src/backend/services/activity_service.py — publish to stream instead of direct broadcast
• src/frontend/src/utils/websocket.js — send last-event-ID on reconnect

Dependencies

None — Redis already running.

[vybe]

Promoted to next priority (2026-04-20)

Per orchestration simplification re-plan — this is the keystone that unlocks Tier 2.5:

• Dissolves bug: long-running backend→agent task calls dropped when uvicorn worker recycles #408 (long-running backend→agent HTTP call gets replaced by push completion; no FaF+polling layer needed)
• Unblocks CAPACITY-CONSOLIDATE (state transitions move to event consumer, which simplifies the drain logic before the three queue classes are merged)
• Unblocks CLEANUP-COLLAPSE (agent becomes authoritative for "is this running?"; most of the 9-path watchdog pyramid retires after a 2-week soak)

See docs/planning/ORCHESTRATION_RELIABILITY_2026-04.md — Tier 2.5.

Scope reminder

• XADD / XREAD with reconnect replay
• Frontend lastEventId support
• Stream trim via MAXLEN ~10000
• WebSocket event shape unchanged (additive)
• Push ships alongside existing HTTP path initially — opt-in per agent until proven

After landing

Gate CLEANUP-COLLAPSE on "#306 in prod ≥2 weeks, zero orphan recoveries observed."