RemoteDelegation.start resumes every open remote task at once, and each
follower read its parent's whole subagent list to find one task, then held
it across a wait of up to ten minutes. A parent with 1,441 tasks (35 MB of
payload) and 122 open ones made every restart decode ~170k records and hold
8 GB, so the server ran out of heap seconds after starting and crash-looped.

getThreadRecords now takes a subagentIds filter. Followers, and task_status
waits that re-read on every parent subagent.updated, read just their task.
Against a copy of the affected database the startup fan-out drops from 8.2 GB
retained over 21.6 s to 8 MB over 61 ms.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>