Design review: retry AI jobs without creating duplicate drafts

How would you design a WordPress AI job so an interrupted run can be inspected and retried? This is an architecture discussion; please label proposed behavior and observed behavior separately.

For a design review, describe:

• How you identify one editorial job across attempts.
• Which job states you record and when they change.
• How a generated draft is linked to the job.
• What happens if saving succeeds but the worker does not receive confirmation.
• How two workers are prevented from processing the same job together.
• Which failures can be retried and which require manual review.

Use a timeline or small pseudocode example if it helps. Include a recovery path for an uncertain outcome instead of assuming every timeout means nothing was saved.

Share one failure scenario your design handles and one it still needs to handle. Replies can walk through that scenario step by step.