fix: tolerate staging disappearance during snapshot cleanup

[danny-avila]

Summary

Snapshot lifecycle inventory can fail with ENOENT when a concurrent publisher removes or renames a staging directory after inventory reads its name. This is inherited from #289 and was found during ClickHouse/ai#4173.

Evidence

• packages/code/src/snapshot-lifecycle.ts calls lstat(directory) in manifest() without handling disappearance.
• publishDependencySnapshot() creates staging outside the maintenance lock and removes it in finally outside that lock. Publishing also renames staging into the final entry.
• Prune inventories names before attempting the owner's key lock, so it can observe the staging name and then lstat it after the publisher removes it.
• This propagates through preparation because only SnapshotBudgetFullError is swallowed, even though the checkout is valid.

Expected fix and regression coverage

Treat disappeared entries/manifests as cache misses during inventory and revalidation, preserving failures for real permission or I/O errors. Add a deterministic test removing/renaming staging after inventory but before manifest inspection; prune and preparation should continue safely without deleting active data.