Surfaced during pear#411 E2E (2026-07-17): the operator's default relaycast workspace roster contains ~86 registered fleet nodes, the vast majority leaked offline test brokers (term-fidelity-*, various experiment instances). Consequences: node pickers and placement eligibility scans iterate dead entries; roster reads are noisy; leaked names accumulate unboundedly. Wanted: (a) a sweep now (deregister offline nodes with no heartbeat for N days — needs a safe CLI/API path), (b) prevention — test harnesses should prefer hermetic local workspaces (pear#411's E2E now does) and/or brokers should support ephemeral node registration that auto-expires. Related: the #411 E2E script's failsafe refuses to run against any workspace containing the live 'pear' node.
Surfaced during pear#411 E2E (2026-07-17): the operator's default relaycast workspace roster contains ~86 registered fleet nodes, the vast majority leaked offline test brokers (term-fidelity-*, various experiment instances). Consequences: node pickers and placement eligibility scans iterate dead entries; roster reads are noisy; leaked names accumulate unboundedly. Wanted: (a) a sweep now (deregister offline nodes with no heartbeat for N days — needs a safe CLI/API path), (b) prevention — test harnesses should prefer hermetic local workspaces (pear#411's E2E now does) and/or brokers should support ephemeral node registration that auto-expires. Related: the #411 E2E script's failsafe refuses to run against any workspace containing the live 'pear' node.