Summary

• Add eidos snapshot + validate steps to training workflow (parity with inference)
• Add conformance evidence collection to training workflow
• Stagger cron schedules to reduce runner contention (T4 :00, inference :15, training :30)
• Normalize branch filter to pull-request/[0-9]+ in T4 smoke test (was pull-request/*)
• Pass unique artifact_name_prefix to gpu-test-cleanup from each H100 workflow
• Generalize gpu-test-cleanup: replace hardcoded gpu-smoke-test pod reference with non-running pods listing
• Add CNCF AI conformance validations to inference workflow:
  • Inference Gateway (fix: add actions:read permission for artifact download #6): verify GatewayClass accepted, Gateway programmed, inference extension CRDs present
  • Accelerator & AI Service Metrics ([Feature]: Configure copy-pr-bot #4/fix: use workflow_run for PR coverage comments on fork PRs #5): verify DCGM Exporter metrics endpoint, Prometheus scraping GPU metrics, custom metrics API (prometheus-adapter)
  • Secure Accelerator Access (fix: rename prometheus component to kube-prometheus-stack #3): verify GPU access is DRA-mediated (no hostPath, no device plugin limits), container security (no privilege escalation), only allocated GPU visible

Test plan

• Verify training workflow runs snapshot/validate successfully on H100 x2 runner
• Verify conformance evidence collection runs in training workflow
• Verify inference gateway validation passes (GatewayClass + Gateway)
• Verify DCGM metrics are scraped by Prometheus and custom metrics API is available
• Verify secure accelerator access checks pass (DRA-only, no hostPath)
• Verify T4 smoke test is unaffected