One dead publish target no longer starves the live ones

When a node repository bakes its NodeType assemblies, the sealed result is published to every portal's storage share listed for the fleet, one after the other. That loop stopped at the first target it could not read: a share that had been deleted with a torn-down instance still appeared in the list, its marker probe answered nothing, and the script — correctly refusing to treat an unreadable marker as an absent one — exited there. Whether the targets after it were published depended on where in the list the dead entry sat; for three days every bake sealed the two live shares and then went red on the third, which read as a broken bake rather than a stale list.

Every target is now published in isolation. A target that cannot be read fails on its own, the remaining targets are still published and sealed, and the run ends red naming exactly the targets it did not reach. The refusals themselves are unchanged — an unreadable marker is still not an absent one — only their blast radius shrank from the whole publication to one target. A target that no longer exists belongs out of the list, and the message says so.