Harden vGPU release ordering in the instance lifecycle

[yummybomb]

Summary

Layer 2 of the vendor VFIO vGPU stack (#366 ← this ← #363 ← #364 ← #321). Where #366 is a behavior-preserving refactor, this layer changes lifecycle semantics — release ordering and retention — still on mdev-era machinery, before the vendor VFIO backend lands above:

• Delete releases the vGPU before other teardown — a successful release persists the cleared assignment before later teardown. A failed release keeps the pre-refactor contract: it is logged and the delete continues, since the release guards never destroy a device they cannot prove is unowned and a skipped release is recovered by startup reconciliation. Stopped instances retry a previously failed release.
• Stop keeps its no-op contract — a failed retained-release on stop logs and retains the assignment metadata instead of failing the stop; assignments are passed to DestroyVGPU as a struct.
• Forks don't inherit retained assignments — a retained (failed-release) vGPU assignment is stripped from fork metadata so a fork can't release its parent's device.
• Start persists a stale release immediately — releasing a stale assignment before start writes the cleared metadata at once, so a later start failure can't leave metadata pointing at a device that's already gone.

Testing

• go build ./..., go vet clean
• go test -race targeted lib/instances lifecycle suites pass

---

Note

Medium Risk
Changes instance lifecycle semantics for GPU teardown, restart-policy blocking on delete, and snapshot restore metadata—incorrect ordering could leak vGPU slots or double-release, though failures remain best-effort with reconciliation.

Overview
Hardens vGPU assignment handling across delete, stop, start, fork, and snapshot restore so host GPU state and on-disk metadata stay consistent when release fails or clones are created.

Delete now blocks restart policy before teardown (with a metadata reload so the block is not overwritten), and releases vGPU immediately after the VMM is gone—before network, devices, or volumes. A successful release is persisted; failures are logged and teardown still continues.

Stop no longer clears assignment metadata when DestroyVGPU fails; the assignment is retained for retry. Stopping an already-stopped instance retries release via releaseRetainedVGPULocked without breaking the no-op contract.

Start releases any stale retained assignment up front and saves cleared metadata before allocating a new vGPU, so a later start failure cannot leave metadata pointing at an already-destroyed device.

Forks (instance and snapshot) strip device path/UUID from fork metadata while keeping GPUProfile; restore keeps the instance’s current vGPU fields instead of resurrecting values embedded in the snapshot.

Lifecycle tests cover ordering, fork isolation, restore behavior, and best-effort delete on release failure.

^{Reviewed by Cursor Bugbot for commit 0fccdc8. Bugbot is set up for automated code reviews on this repo. Configure here.}

[yummybomb]

Trimmed over-engineering after an architectural review against the parent task (host kernel 6.8 / Ubuntu 24.04 support):

• 1c1f52e — dropped the HasMdev/IsAllocated() compatibility shim. VirtualFunction is internal and never serialized in an API response or persisted, so a single Allocated field is enough.

Kept as-is:

• the delete/stop behavior change (fail and retain assignment metadata when release fails) — orphan reconciliation only runs at startup, so failing the delete is the correct fail-closed behavior;
• the darwin DestroyVGPU unknown-framework error — it keeps the fail-closed rollback semantics identical on macOS, which the lib/instances tests exercise in CI.

Tests: go test -race ./lib/devices, ./lib/resources, ./lib/hypervisor/qemu, ./lib/hypervisor/cloudhypervisor, and the vGPU-targeted ./lib/instances tests all pass; the remaining lib/instances E2E failures require KVM/images and fail identically on an unmodified checkout.