Every worker-list recipe now passes --run <run_id>. Unscoped, a real profile
returned 2103 rows and none of the live workers, so the stall recipe the kernel
prescribes could not find them.
Reword the disagreement rule: the fleet verdict still decides, except against an
unverifiable reason that names a client-side gap (missing_status,
host_unavailable), where a worker-show verdict from the execution host is the
better evidence. Note that a peer lacking the fleet-snapshot capability also
reports host_unavailable on the row while the host warning names
capability_unsupported. unverifiable still authorizes nothing.