agentmother heartbeat loop dies while A2A server stays up (passive agent goes silently offline)

[NeilJo-GY]

Scope note

The fix lives in the agentmother agent codebase (agent-side), NOT in ACN.
Tracked here only because this is the team's active issue tracker. Move to the
agentmother repo if/when that tracker exists.

Symptom

A paid Store order's store.order_paid message did not trigger agentmother's
/a2a -> poll_orders() in real time. Investigation showed:

• The agent's HTTP A2A server was up and reachable (http://43.160.236.79:80/a2a).
• But ACN considered the agent offline: the Redis alive key had expired
  (TTL 60min, renewed on each 15min heartbeat), so the message was parked in the
  inbox instead of being pushed.

Conclusion: the heartbeat loop stopped for >60min while the HTTP server kept
running — classic "async background task swallowed an exception and exited
silently".

Why it matters

Even with the ACN-side fix (ADR-0012 "reachable == online", commit 5683143),
which now probes reachable endpoints regardless of the alive key, a robust
agent should keep its own liveness signal healthy. A dead heartbeat also
corrupts online/offline stats and any logic that reads agent status.

Action

• Find why agentmother's heartbeat loop terminated (>60min gap)
• Make the heartbeat task supervised/restarting: never let an unhandled
  exception kill it silently; log + auto-restart with backoff
• Consider an implicit-heartbeat fallback: renew liveness on any successful
  inbound A2A handling, not only via the dedicated heartbeat call

Unblock the stuck order

Order 53cce555-6dae-4d1c-b62e-eaf19402de83 is paid and sitting in the inbox;
bringing the agent online (or its cron inbox-pull) will process it. Not lost.

[NeilJo-GY]

ACN-side is already closed; remaining work is agent-side (confirms ready-for-human).

Re-reviewed against current main. The two things ACN can do for this are both
already in place:

1. Core symptom (heartbeat expired → message parked in inbox instead of
   pushed): fixed by "reachable == online" (5683143) — route() no longer
   gates on the Redis alive key; a reachable endpoint is pushed regardless.
2. Implicit heartbeat (the issue's "renew liveness on successful inbound
   handling" suggestion): already implemented — a confirmed delivery calls
   touch_alive(to_agent) (message_router.py), so an agent producing/receiving
   traffic stays "online" without the dedicated heartbeat cron, fixing the
   online/offline stat drift.

The three open action items (find why the heartbeat loop died, make it
supervised/auto-restarting, agent-side implicit-heartbeat fallback) all live in
the agentmother codebase. Nothing further actionable on the ACN side.