Job status + logs (?stream=&offset=&tail=&attempt=&links=)

Show one job by id: status, GPU, attempts, and while queued its position and a wait estimate (moves both ways; a forecast, never a countdown; queue.wait_reason names the gate; null beside "queue_order" = no signal yet). Hours queued is normal (the spot holds for queue_ttl_minutes): pace polling minutes apart off the estimate; failure reason and exit code arrive once terminal. While building, build carries live progress (phase, byte/layer counts, updated_at): multi-GB images download for minutes, updated_at advancing means it is moving, and byte counts hold still past fetching by design - read the phase, and stream: "build" for what the build is doing. While starting, startup is the same shape for the run's own image pull (absent fields = no signal yet, never stalled). gman://docs/llms.txt covers attempt_history (end_cause, end_reason) and mid-run telemetry (checkpoint_capture, preview, samples). A completed attempt carries result (the $GMN_RESULT_PATH JSON, verbatim; result_error names the failure) and artifact (the $GMN_OUTPUT_DIR capture, a presigned download on the object store hostname, currently fly.storage.tigris.dev - an egress allowlist must permit it; a severed attempt uploads nothing); with output_to declared it instead carries output: {connection, uri} - the bytes land in your own store. Pass stream ("build" or "run") for incremental logs (offset to advance, tail: true for the last 4 KiB, attempt for an earlier attempt). A run-log slice answers the poll question: attempt_state is "not_started" (no run log YET; poll the JOB), "live", "complete", "never_ran" (no container, NO log ever; an empty slice is final), or "unknown" (our read failed; read again), and eof: true means stop; note names an unreadable range (skipped at next_offset, never spliced) or a job that ran no build. Logs live 30 days; status checks never touch a worker. links: false omits download links. Task sets add tasks. Poll it after submit_job.