bug: make subagent completion delivery durable under busy-lane and restart conditions

[100yenadmin]

Problem

Subagent results are being lost or delayed because completion delivery depends on synchronous direct announce into a busy parent lane. Restart/drain windows and local loopback tick timeouts make this worse.

Findings so far

• dominant failure mode is announce-path design under congestion, not just transport instability
• drain/restart windows reject announces outright
• retries amplify congestion
• there is no durable spool/inbox protecting results after retry exhaustion

Goal

Make completion durable first, delivery second.

Proposed direction

• persist failed completion payloads to a durable queue/inbox
• short-circuit direct announces during drain/restart states
• batch/opportunistic later delivery when the requester lane is available
• instrument effective timeout, queue depth, and event-loop lag

Acceptance criteria

• no completed subagent result can be silently lost after retry exhaustion
• restart/draining states persist instead of re-hammering direct announce
• blocked parent lanes delay delivery, but do not destroy result visibility

Notes

This should be traced in code, not just from logs. GitHub-native fix path only.