fix: keep subagent cleanup running with large retained tasks (#201960)

Archiving a backlog of completed subagents with large retained tasks could repeatedly kill the state-read worker. Deletion-only mutations unnecessarily restored every selected full payload before planning removal, so every sweep retried the same growing payload set.

Plan deletions from resident lifecycle metadata under the existing FIFO reservation and publication owner, and reject non-null postimages on that path. Ordinary updates still restore full payloads. Process DELETE RETURNING in groups of 16 inside the existing transaction so removal receipts remain atomic without collecting the entire deleted payload set in memory.

A real retained state-read worker limited to 64 MiB fails on the base with the channel-closed warning when archiving 108 synthetic 384 KiB tasks. With this fix, consecutive ticks each archive all 108 rows and durable reads confirm deletion. The focused suites passed 83 tests; the expanded cross-batch rollback and capacity rerun passed 16 tests. Net test-line delta is zero. No schema, configuration, retention, or stored-payload change.

Follow-up for Peter: evaluate whether retained runs need to store the full task text per row.