Right-size event_log telemetry retention: 7d raw is ~800MB steady state, gates_evaluated JSON is a third of it

event_log grew to 577MB / 1.2M rows in ONE WEEK before v4.63 added the hourly retention prune, on a DB read every 30s. The prune keeps 7 days of the three per-tick telemetry events (observed/gates_evaluated/decided) plus tick scaffolding, so steady state is ~800MB and will not shrink further without either a shorter RETENTION_DAYS or not writing gates_evaluated's full JSON payload every tick per agent (272MB of the total was that one column). Decide: is a 7-day raw telemetry window actually needed for anything, or would a daily rollup plus 48h of raw serve every query we run? Also note VACUUM is manual by design - it is not safe to run inside a 30s-cadence tick - so the file only shrinks when someone runs it.