Command status + log (offset-polling)

Status + incremental output of a detached command: {status, exit_code?, output, next_offset, truncated, result?, note?}. result is the JSON the command left at $GMN_RESULT_PATH, verbatim (result_error names the honest failure: invalid JSON, or over 32 KiB). output is bytes of the log from offset (default 0); pass next_offset back verbatim to stream; truncated:true means more is available right now. Status checks NEVER wake a stopped node: while the node is stopped only the final tail captured at exit is served - a read below the tail start returns empty output with next_offset jumped forward and a note naming the gap (the full log survives on the disk; wake the node to read it all). Polling this does not keep the node awake or reset its idle grace. A supervised command (restart:"on-failure") also reports {restart, max_restarts, restarts, last_exit} - flapping is visible here without reading logs; status "crash_loop" means the restart budget exhausted on a nonzero exit. oom_kills (with a note) means the kernel's OOM killer fired in the node's container during the run - a bare "Killed" in the log plus exit 137 is the container's memory ceiling, not a bug in your script; get_node's resources carries the budget. Status "lost" means no exit status was recorded - the container went away before the exit reached us, so there is no exit_code and no result; note says what survived (the volume's files and the log, unless the node was lost with its host); on a live node this call also re-reads the command's own exit record from the volume and closes a lost row with it. A run whose verdict you need must reach an exit_code here BEFORE you stop the node. A command that NEVER STARTED (an on_wake boot command whose launch failed or lapsed) has no log at all: output is the reason it never ran, and there is no exit_code because nothing ever exited.