The runner has Tier 1 (an ephemeral buildkit scale on an existing fleet node with headroom). Tier 2 (create a net-new on-demand hcloud node when no fleet node qualifies) is NOT wired for the exec transport: an on-demand node is not cluster-enrolled, so the hydracluster exec channel cannot reach it - it would need direct SSH provisioning + a temporary exec path, then teardown. Today a placement refusal fails the build. Fine while the arm64 fleet has headroom; a scaling gap otherwise. The internal/hcloudprov + placement Tier-2 scaffolding exists; finish the dispatch-over-SSH path and the always-deferred teardown. Related: #508.