Aggregate DMARC and TLS reports are queued with their first attempt at a random time up to three hours later, so reports from every server don't reach receivers at once. Delivery time was counted from when a message was queued, so a report held for hours on purpose and sent on its first try showed up as hours of delivery time: the console's delivery-time graph spiked and its average read 18 minutes against 2 seconds per attempt. send_report now records the hold in the message's flags, in the top 14 bits (50-63), which no flag uses, and the delivery-completed event's elapsed time counts from the end of the hold. That event feeds the delivery-time histogram, the stored metric the dashboard and scheduled reports read, and delivery history. The hold itself is unchanged. Spare bits rather than a new field, so the stored queue format doesn't change: a node that doesn't know them ignores them and measures as before, and the console's flag list only reads the bits it knows. Messages already queued carry no hold and are measured as before. Reported on Discord (#inbuxa). Fixes #517