docs(adr): ADR-0001 — SSE transport tier-2 evaluation (status-quo)

[howard-eridani]

Summary

• Research spike: evaluate whether SSE cursor+replay (shipped May 21) is sufficient, or if we need NATS JetStream per-agent consumers or Redis Streams as tier-2 transport.
• 21-day prod observation window (2026-05-21 → 2026-06-11).
• Decision: C — status-quo (cursor+replay on existing JetStream+Postgres stack is sufficient).

Key Findings

• NATS JetStream MESH_EVENTS already the durable write-path backbone (running since 2026-03-13): pg-writer consumer shows 0 redeliveries, 0 pending across 13,228 events.
• agent_events table retention: 7 days (critical events: task.assigned, task.created, task.mentioned) / 24h (others) — covers all reconnect scenarios within normal operating patterns.
• 0 lost-event incidents, 0 stale-redispatch triggers in 21-day window.
• Option A (per-agent JetStream consumers for SSE fan-out) is a valid future upgrade at 50+ agents; not warranted now.
• Option B (Redis Streams) would be redundant — JetStream already provides the same guarantees.

Follow-up (non-blocking)

• F1: sse_cursor_410_total Prometheus counter (closes telemetry gap)
• F2: sse_replay_events_total counter
• Both are separate small tasks for Linus, not part of this ADR.

Test plan

• ADR reviewed by Pavel
• If approved → task fc2acff3 closed (no implementation needed per acceptance criteria)
• F1/F2 counters tracked as separate follow-up tasks