List cluster nodes.

Runtime's per-process boot id, forwarded from the node's /capacity. A change between heartbeats means the serve restarted (its warm pool wiped), which the control uses to prune the node's now-stale pool records. Absent on an older node/runtime that doesn't report it.

PEM-encoded certificate signing request, sent only at registration. When present, the control plane signs it as the node's mTLS server cert (overriding the CSR's subject/SANs with the validated node id/addresses) and returns the cert in RegisterResponse. The node's private key never leaves it. Absent ⇒ control plane has no node-CA configured, or this is an older node; the registration still succeeds (cert fields stay empty).

Set (RFC3339) when the control plane soft-cordons this node for DISK pressure — its heartbeat-reported host-free disk dropped below the cordon threshold, so it stops receiving NEW placements. SEPARATE from the data-plane cordon: the node stays status = "ready" and is never reaped; its running machines are untouched. Cleared when free disk recovers past the (higher, hysteretic) un-cordon threshold. None = not disk-cordoned.

Sum of the (thin/nominal) local-SSD footprints reserved by the machines placed on this node — the disk-overcommit ledger. Unlike available_disk_gb (overwritten by every heartbeat with the host's live free space), this is a reserve/release decrement ledger the scheduler gates admission on, so a node can't be packed past total_disk_gb * ratio. Reconciled by the capacity self-heal. Default 0 (legacy rows / never reserved).

The runtime this node speaks (e.g. "smolvm"). The control plane resolves a driver from this name; smolvm_address is that runtime's endpoint. Defaults to "smolvm" so a node that doesn't report a runtime (the only runtime today) is handled transparently.

The runtime endpoint (HTTP address) on this node.

The smolvm engine version this node's runtime reports on /health, recorded by the control plane. Read-only: a node never sets it. A checkpoint is only restored on a node at the capturing engine's version or newer. Absent until the runtime has been probed.

Real CPU usage as fractional CPUs (e.g., 2.5 = 2.5 CPUs worth of load). Reported by the node from hypervisor/process stats. Separate from available_cpus (which is allocation-based bookkeeping).

Real disk space consumed by VM overlay/storage files in GB. Reported from filesystem stat of VM disk images. Separate from available_disk_gb (which reflects the entire host filesystem).

Real memory usage in MB. Reported by the node from RSS of VM processes. Separate from available_memory_mb (which is allocation-based bookkeeping).