fix(vm): reclaim cancelled image preparation after worker exit

[shiju-nv]

Summary

A timeout or cancellation can abort a sandbox's async preparation task while blocking filesystem work continues and temporary files remain. Run image and overlay preparation in an owned worker process. On cancellation, stop its process group and wait for cleanup before reclaiming files. Recover inactive attempts after driver restart.

Related Issue

Fixes #3953, accepted by a maintainer on 2026-09-30. Part of #3955. Depends on the workload-identity change in #4036.

Changes

• Register each worker before yielding. Stop/delete join the cancelled task, terminate its process group, and wait for owned processes to exit. Inherited file locks protect files until descendants stop; uncertain cleanup retains staging and sandbox state.
• Serialize shared cache publication and atomically publish complete disks, including writable-overlay retries. Reclaim marked inactive attempts at startup while preserving active attempts, committed caches and unmarked legacy staging.
• Carry user/group selectors across the worker boundary and validate them against the persisted overlay owner before writes. Preserve successful worker status when stdout closes before exit, including Darwin's exited-process-group EPERM case.

Testing

• mise run pre-commit passes locally. Focused local checks and the linked Branch Checks passed.
• Unit tests added/updated
• E2E tests added/updated (if applicable). Actual-worker integration tests passed; physical-VM testing has not been rerun for this candidate.

Checklist

• Follows Conventional Commits
• Commits are signed off (DCO). Both MG03 and MG06 signatures and DCO were verified.
• Architecture docs updated (if applicable)

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

[github-actions]

🌿 Preview your docs: https://nvidia-preview-pr-4039.docs.buildwithfern.com/openshell