Make stacked content pruning reference-aware

[yzhuang-oai]

Split from #1307. Adds reference-aware content-store pruning and GC safety for stacked disk content, including race-focused regression tests.

Validation: full Swift test suite passes locally.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: dddf12157e

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 9001d8c340

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: af825bd3da

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

💡 Codex Review

tart/Sources/tart/VMStorageOCI.swift

Line 343 in dfd52fa

let contentURL = try contentStore.contentURL(for: contentDigest)

Validate content digests before assigning owners

When a structurally recognizable stacked cache record contains a malformed whole-file digest such as sha256:bad, diskContentDigests() returns it successfully because the manifest parser only checks that the annotation exists. This call then throws invalidContentDigest, causing VMStorageOCI.prunables()—and therefore manual and automatic cache pruning—to fail entirely because of one corrupted record, despite the preceding comment saying malformed records should not block pruning. Skip invalid digests along with other malformed manifest data.

---

tart/Sources/tart/Commands/Prune.swift

Lines 195 to 196 in dfd52fa

	try prunable.delete()
	currentCacheUsedBytes = try prunables().map { try $0.allocatedSizeBytes() }.reduce(0, +)

Measure actual reclaimed bytes after reference transfers

When a concurrent stacked clone publishes a local manifest while automatic pruning is deleting its cached source record, the shared content moves from the record's attributed size into nonCacheContentDigests; deletion correctly preserves the file, but this recomputation drops its bytes from the prunable total anyway. The loop can consequently conclude that gigabytes were reclaimed even though only the small cache record was removed, leaving the pull that requested space to fail with ENOSPC. Base progress on actual freed storage rather than the changing set of eligible prunables.

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".