[SaladDay]

Stacked on #815.

A microsandbox suspend exported the native snapshot to a zstd archive, fsynced and hashed it, then deleted the native snapshot; every resume hashed and unpacked the archive again. On mx that export was 2.73 s of a 5.85 s suspend and about 0.9 s of each resume.

Now the native snapshot is the checkpoint:

• Suspend creates the full snapshot directly in the private checkpoint root (SnapshotCreateOptions.DestDir, group s-<operation>). The SDK materializes the checkpoint closure there (hard links on the same filesystem, otherwise a synced copy) and returns it durable. The adapter verifies it once, writes the ownership receipt, stops the source, then reports SourceStopped.
• Resume verifies the snapshot once and restores from it directly. A node that mounts the same checkpoint root restores it the same way, so checkpoint transfer keeps working.
• DeleteSnapshot writes the tombstone, then removes everything else in the operation directory without opening the snapshot, so a damaged snapshot or an interrupted capture's staging directory never blocks cleanup.
• Archive export, import, hashing and their receipt fields are deleted.

Cost: a retained checkpoint takes about 0.5 GB instead of about 0.15 GB, and a shared checkpoint root receives a full copy at suspend.

Checks: helper tests with the bundled SDK (scripts/build-microsandbox-provider.py test), go test for internal/sandbox/microsandbox, Core go vet ./internal/sandbox/..., translation check. Real-cluster timings on mx are in progress.

---

^{Need help on this PR? Tag @codesmith-bot with what you need.}