Compute orchestrationScheduling

Queue and object layout

Trace a job through queue, running, status, completed or failed, output, capacity, and registry objects.

Queue and object layout — contract and behavior

Object prefixes encode lifecycle ownership and provide a durable ledger independent of any scheduler process.
Claim transitions prevent two consumers from treating one queue record as theirs.
Per-job status and output are written separately from terminal lifecycle records so large artifacts do not bloat scheduling scans.
Capacity/<consumer-id>.json is refreshed every poll cycle and expires operationally when a consumer stops broadcasting.