fix(make): prefer existing local artifacts

[MeteorsLiu]

Prefer artifacts already installed in the local LLAR workspace and persist public remote hits for later commands.

The implementation includes:

• Changes readThroughCache.Get to return local cache hits before querying the public Kodo cache.
• Writes successful remote hits through the local cache so the restored artifact receives a .cache.json entry.
• Surfaces local write-through failures and keeps source-build Put operations scoped to the local cache.
• Covers local hits and errors, remote hits and errors, complete misses, write-through failures, .cache.json persistence, and repeated reads that skip remote access.
• Validates the affected paths with go test -ldflags="-checklinkname=0" ./cmd/llar/internal ./internal/build/cache ./internal/crosscompile/...; the full local suite previously reported the unrelated macOS zlib v1.2.11 build failure in TestE2E_RealLibpngBuild.

This prevents existing artifacts from being overwritten and avoids downloading the same public artifact again on later llar make or llar test runs.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[fennoai]

Review: prefer local cache before remote in readThroughCache

The core change is small and correct. Swapping readThroughCache.Get to read the local workspace cache first and fall back to remote short-circuits a per-module HTTP round-trip on incremental builds and lets locally built artifacts take precedence for a pinned version+matrix. This is consistent with Put writing only to the local cache. The refactored table-driven test in make_test.go is clear and locks in the ordering plus the per-branch call counts (gets/remoteGets). Nice.

One documentation fix worth making, plus two design notes for the record.

Should fix

• Stale comment contradicts the new read order — cmd/llar/internal/make.go:208-209. The comment at the readThroughCache construction site still describes the old remote-first behavior:

  // Reuse published artifacts from the public Kodo store when available, and
  // fall back to the local workspace cache and source builds otherwise.

  After this PR, Get reads local first (lines 54-60), matching the updated struct doc at lines 46-48 and the new tests. This inline comment now states the ordering backwards and will mislead the next reader. Suggested rewrite:

  // Reuse artifacts from the local workspace cache when available, and fall
  // back to published artifacts in the public Kodo store (and source builds)
  // otherwise.

Design notes (non-blocking — confirm this matches intent)

• Local-first shadows republished/security-patched remote artifacts. Cache keys are keyed only on path/version/matrix. Once a local .cache.json entry exists for a key, the remote artifact for that same key is never consulted again. If a corrected artifact is ever republished under an identical version+matrix, users with a populated workspace cache won't pick it up. If artifacts under a fixed version are considered immutable, this is fine — just confirming it's intended.

• Local hit trusts installDir as-is without re-materializing or verifying it. On a local hit, localCache.Get (internal/build/cache.go:90-100) returns only Metadata and the builder trusts the on-disk install dir. Unlike the remote path, there's no checksum re-verification, and the dir isn't re-unpacked. If the .cache.json entry survives but its installDir is deleted/partially removed, the local-first path reports a hit with a stale/empty output instead of falling back to remote. Worth confirming that the metadata entry and its installDir are always written and invalidated together (or validating installDir existence on a local hit).

Additional findings

• cmd/llar/internal/make.go:209: [P2] Stale comment describes old remote-first cache order: This comment still describes the pre-PR remote-first behavior, but readThroughCache.Get now reads local first and falls back to remote (matching the updated struct doc at lines 46-48 and the new tests). The ordering is stated backwards. Suggested rewrite:

[luoliwoshang]

》 LLARD Docker Image / build (pull_request)
LLARD Docker Image / build (pull_request)Failing after 10m

[luoliwoshang]

LGTM