When a run fails and a queued run targets the same provider, the
orchestrator holds the thread's queue until a user message or the UI's
queue.resume releases it. An MCP orchestrator could neither see that hold
(t3_queue_list only returned ids and text) nor clear it, so a dispatcher
thread driven by scheduled tasks and webhooks sat with its queue stuck.

t3_queue_list now reports a thread-level `held` flag and `held` per item,
and the new t3_queue_resume tool dispatches the existing queue.resume
command, surfacing the usage-limit rejection as a clear error.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>