The shipped scheduler brief said to hand over the prompt and use the same model,
and never said what to start. So an agent following it had to guess a subagent
type, and the obvious wrong guess is the role's own name — there is no
`secretary` type, and asking for one fails with `Agent type not found`.

A peer chased that to the bottom tonight: the binding named an agent that cannot
be started, `orchestrate on` accepted it because the ROLE FILE exists, and every
secretary that appeared to work was a `general-purpose` agent standing in. Nothing
downstream could have been right, because the thing dispatched was never the thing
named.

The brief now says it, and the refusal at binding time says it too, so nobody has
to learn it from a failed dispatch. And it names the repair NOT to make: starting
something else under the role's name clears the error and writes a false
attribution into the record.

That last part is a probe's own conduct, and it is the better half of this. Told
six times to dispatch a secretary, it tried, got `not found`, and declined to
substitute — "that would make the trigger record a filing by an agent the binding
never named" — choosing to keep firing rather than lie in the record. Every
secretary that seemed to work did what it refused to do.