One node: state, rate, queue position

Show one node by name: status, image, the effective rate, its region, and, while queued or provisioning, the queue position and wait estimates (estimated_wait_seconds; estimated_ready_seconds and its p90; pace polling off the p90; null = no history; your place is held). Poll until running, then run_command; while queued, arrival_note says what to do now. While placed, resources is the container's cgroup budget (cpu_limit, memory_limit_bytes) - size work to those, not to os.cpu_count()/free, which report the HOST - and environment the CUDA-compile facts (image_nvcc, image_torch, driver_version, cuda_major_max, compute_capability, gpu_graphics, gpu_graphics_enabled: GPU graphics is off by default, on per organization on request via file_ticket) - node_doctor checks the pairing. gpu_memory is per-GPU used/free VRAM: read free_mib_min before co-locating work. A network member shows network (dns_name, region/zone, data_plane: up, converging, not_placed, or unavailable). A ready node with no session reports "running (idle)" plus idle_stops_at - when it auto-stops unless a run_command lands first; hold_node keeps it awake (shown as hold). A silent host (execs failing "host isn't responding") shows as host {responding: false, silent_seconds}; absent = it answers. Status checks never wake a stopped node. A stopped node carries stopped {reason, at, note, ran_seconds (billed length)} plus wake_estimate (null if it cannot start): what a wake now would be quoted (queued_wakes, ready estimates; a reading, not a reservation; assumes_contested prices contention). durable_as_of is the newest instant of the disk replicated off its machine (storage tier: docs). A retiring machine shows machine_retires_at and machine_retirement_note about 45 minutes ahead, for a parked disk too; poll for it on long runs. A failed create answers status "failed" with the real error for an hour; re-creating the name starts fresh. A shell can poll GET /preview/nodes/{name} (docs, HTTP API).