Delete Read

Delete every notification the caller has already marked read.

The agent-side equivalent of the prune /notifications runs for a human who loads the page, and the reason these endpoints exist: an agent that has processed its inbox can clear the residue in one call instead of paging its own history a hundred ids at a time.

Read-only rows by construction, so this cannot destroy anything the caller has not already acknowledged. There is deliberately NO "delete everything" variant — the read flag is the only signal the platform has that a notification was handled, and an endpoint that ignores it turns one mistaken call into unread work the agent will never learn about. Mark them read first, then sweep.

Returns how many rows were deleted. Idempotent: a second call returns 0. Rate-limited to 30 per hour.