Extend the existing IR, generated Python workloads, independent oracle, regression corpus, resource accounting, and shrinker for the five-node normal/recovery/survivor campaign. Retain exact fixture identities across attempts and persist coverage, failure, timing, and cleanup evidence.
Add raw Docker/SSH redeployment fixtures, bounded campaign execution, durable paid-run cleanup state, ordered gate attestation, and scripted-provider safety scenarios. Include the offline oracle reproduction utility for focused diagnosis. Preserve paid access guards; no paid acceptance is claimed.
Review verification: 169 harness library tests and 4 ordered-attestation tests passed. The retained ordered run passed Gate A and the full local campaign, then exceeded the complete workflow timing ceiling before later safety checks.
Stream bounded provider logs and surface provisioning phases through telemetry and the dashboard. Preserve offer-search criteria, use live relay endpoints for remote bootstrap, and add GPU smoke coverage. Exercise orchestrator crash recovery and clean completed checkpoints without dropping active work.
Add a remote-node file picker that uploads one Python file through the data namespace, materializes it over Iroh, and launches it with contextual-process bootstrap while streaming lifecycle and output events. Clean up execution artifacts, install Python and the swactor wheel in the production node image, and build the worker inside Docker to prevent stale host binaries.
