sweepAccess sweeps each forge in its own goroutine, but nothing pinned
that: the table test asserts coverage, not ordering, so the sweep could
be made sequential again without a failure.
The test stalls the root forge in AcceptInvitations and asserts the
sibling identity is swept while it is still blocked, then that Startup
returns once the stall is released. Verified RED against a sequential
sweepAccess (times out waiting for the sibling) and GREEN against the
goroutine-per-forge implementation.