v0.2.18

• A job the server accepted is never lost again — and a wait timeout no
  longer reads as a failure. These were the two behaviours that made a human
  re-run a generation and pay for it twice (NOL-344 cost 84 credits exactly
  this way).

  Previously, if anything went wrong after a submission had already succeeded,
  the CLI printed a transport error and no job id — so the command looked
  like it had failed before submitting, and re-running was the natural
  response. And a 408 from GET /jobs/{id}/wait, which only means the
  server's long-poll window closed while the job kept running, surfaced as:

  Error: waiting for generation job
  
  Caused by:
      Unexpected Response: Response { url: "…/wait?timeout_seconds=300", status: 408, … }
  

  Nothing had failed, but the word Error: is the strongest possible prompt to
  try again.

  Four endings now deliver the same fact — work is live under this id, follow
  it, do not re-submit:

  still running after 300s — job 60893909-3123-42bd-b04f-ed946b136c0f
    Nothing failed. The server's long-poll window closed while the job was
    still running — the job was not cancelled and is still being worked on.
    It will be billed once, whether or not you keep waiting. Re-running this
    command would start a second job.
      nolgia wait 60893909-…    # keep waiting for it
      nolgia status 60893909-…  # check it once
  

  • gen now prints the job id on stderr the moment the server accepts it,
    before any waiting begins. This is the robust half: an error path can only
    speak if the process lives long enough to reach it, and in the incident it
    did not. The line is already in the operator's scrollback even if the CLI
    is killed outright or the pipe is torn down.
  • Ctrl-C after submission, and any other post-submission failure, report the
    job id and how to follow it instead of a bare error.
  • A duplicate submission — refused by the API with 409 since NOL-344 — is
    rendered as the job that already exists, restating the API's advice as
    commands a shell can actually run (the server says "check it with
    GET /jobs/{id}", which is true and unusable at a prompt).
  • gen audio refusals go through the RFC 7807 handler like image and
    video already did; they previously surfaced as raw Unexpected Response
    debug dumps. nolgia status likewise no longer dumps a raw response on a
    404.

  New exit code 75 (sysexits EX_TEMPFAIL) means "a job is live; do not
  re-run" and is used for all four cases above. It replaces exit 1 only
  there; every other failure still exits 1 with the same Error: text as
  before. This also makes the situation machine-readable for the first time:
  a 408 was previously indistinguishable from a genuine failure, which is
  why nolgia-agent's CLI backend could not implement the "keep polling"
  contract its SDK backend has always had. Under --json, stdout carries a
  document (job_id, outcome, billed_twice, follow_up) while the human
  block goes to stderr, so a program's stdout stays parseable.

• --idempotency-key (also NOLGIA_IDEMPOTENCY_KEY). The API's duplicate
  guard fingerprints the request body, so a deliberate second take of an
  identical prompt is refused too; the documented escape hatch is an
  Idempotency-Key header, which the CLI previously had no way to send — the
  header is accepted by the API but is not declared in the OpenAPI spec, so the
  generated client emits no parameter for it. Passing a fresh key runs an
  identical request again on purpose; reusing one collapses a client's own
  retries into a single job.