One live fleet session to verify what already shipped #89

Tasks

• 1. Capacity truth (provider-capacity-truth 3.1): with one host serving a request,
  peers/opencode agents show it 0/1 slots free and an idle host 1/1; a subagent
  lands on the idle one
• 2. Background default (subagent-background-default 6): a local orchestrator fanning
  out three tasks does not block on the first; all three land on different hosts or peers
• 3. Cloud parent → local subagent (role-placement-policy 4.2): a role with
  placement: "local" run from a cloud model places its subagent on a local host
• 4. Context-aware placement (ctx-aware-subagent-placement 10–12): a subagent prompt
  larger than a small-ctx host's max_safe_ctx skips that host; the placed model's
  limit.context equals the host's current max_safe_ctx
• 5. Delegation (skein-pool 2.5): parent on a full single-slot host + idle Claude Code
  peer → task envelope arrives at the peer, its [peer-task-result] reply lands as the
  task result; repeat with an idle opencode peer on a cloud model
• 6. Eternal loop (loop-eternal-by-default 6.2): plain /loop on a quickly-completing
  prompt keeps going until stopped; --until-done stops on completion
• 7. Persona fan-out (persona-gate-fanout 4.5): /backlog on a real change shows the
  reviewer persona's parts inline