[joecklau]

Summary

Upgrading Docker Sandboxes from v0.46.0 to v0.49.1 caused ./start to fail on inventory and diagnostics CLI changes introduced in v0.48. Long-running pools could also exit when a host-trust lease write timed out during runner readiness. This PR restores startup compatibility and keeps recoverable candidates within the existing quarantine, exact-cleanup, and capacity rules.

• Support legacy and current inventory, policy-log, and diagnostic formats, including workspace-free sandboxes and optional diagnostic presentation fields. Preserve strict identity, policy attribution, freshness, and recovery-census checks.
• Route readiness lease-write timeouts through bounded candidate recovery, while retaining terminal handling for cancellation and permanent trust failures.
• Verify stop, deletion, template removal, and policy outcomes through fresh structured readbacks instead of command prose. Finish receipt-owned staging cleanup after deferred deletion or controller restart before tombstoning the instance.
• Require actual apt/dpkg lock absence before guest readiness, including locks held by unfamiliar processes. Preserve the narrow known-bootstrap termination rules, bounded waiting, and recovery marker; ignore zombies that cannot hold locks.

The changes are split across seven focused commits. The guest lock check requires a rebuilt local template or a newly published prebuilt package containing it. Live Windows relay validation remains outstanding; the macOS smoke does not certify foreign-architecture workloads or other host platforms.

Validation

Validation was performed across the commits and repeated for affected behavior after corrections:

• Full Go suite (go test -timeout 5m ./...), affected image/provider/pool/startup normal and race checks, and wrapper syntax/command-forwarding checks passed.
• Windows amd64 provider test binary compiled successfully; legacy and current policy-log formats passed the relay attribution, freshness, route, and rule-proof regressions.
• On macOS arm64 with sbx v0.49.1, default ./start brought two cached Full-template runners online and idle. Console and guest logs, host-trust lease renewal, and exact shutdown cleanup were checked. A foreign workspace-free sandbox remained untouched by EPAR cleanup.
• A fresh Act local build passed provenance/SBOM generation, exact template import and activation, guest helper checksum verification, runtime admission, GitHub online/idle, and exact shutdown cleanup.
• A disposable Linux sandbox verified natural release of an unknown holder of the real dpkg frontend lock, a 185-second bounded timeout with the recovery marker while preserving the holder, and readiness with an unreaped apt zombie.
• Independent reviews completed with no outstanding correctness blockers. Final live status and sandbox inventory were empty after cleanup.

Checklist

• I kept credentials, private keys, tokens, and machine-specific configuration out of this pull request.
• I added or updated tests where behavior changed.
• I updated relevant documentation.
• For provider or onboarding changes, I followed the Development and Extension Principles and documented and tested every intentional exception.
• I read and followed the contributing guide and code of conduct.