Who actually read from which vault

Reachability (the topology's policy edges) says who could read a vault. This says who did, counted from audit_events.action = 'secret.read' — the durable record every read already writes — grouped by agent and vault. That is what lets a Sankey ribbon's width be a count of real reads rather than a guess.

Agent actors only; a human reading a secret through the dashboard is not an agent flow. Deleted agents and vaults are kept and labelled as such — a read that happened is still a read.