GeezerVison

LLM Description written by Claude Code:claude-opus-5-5

Fixes #1541

Summary

A repo whose .gitattributes hands *.tgz to Git LFS committed every vendored npm-family tarball (and each bun.lockb workspace mirror) as an LFS pointer. Any checkout that doesn't smudge LFS (actions/checkout defaults, no git-lfs, GIT_LFS_SKIP_SMUDGE=1, air-gapped builds) then failed the frozen install, while the vendored scan reported success.

Root cause

All npm-family vendored flavors (npm, pnpm, yarn classic/berry, bun text and bun.lockb, vlt) write .socket/vendor/npm/<uuid>/.gitattributes from one constant, npm_dir::UUID_GITATTRIBUTES. It held * -text, which turns off EOL conversion but leaves a parent filter=lfs rule in force.

Fix

• UUID_GITATTRIBUTES is now * -text -filter -ident -working-tree-encoding: every attribute that lets git store or check out bytes other than the artifact's is unset at the uuid level, which beats any .gitattributes above it.
• The pre-Vendored Bun tarballs go into Git LFS when .gitattributes tracks *.tgz with LFS, so a checkout without LFS fails bun install --frozen-lockfile #1541 * -text file still counts as ours (is_uuid_gitattributes), so a Bun workspace mirror's metadata is still removed on revert. An in-sync vendored re-run already passes through restore_uuid_metadata (via stage_patch_pack → keep_pack_committable) for all six tarball flavors, so existing trees are upgraded on their next run.
• mirror_metadata now checks the file name before reading a sibling.
• CLI_CONTRACT (npm and vlt rows) and docs/ecosystems.md describe the new attributes, plus the git add --renormalize .socket/vendor/npm step for a tarball an older release already committed as a pointer.

Tests (red → green)

Local runs: cargo fmt --all -- --check ok; cargo clippy --workspace --all-features -- -D warnings ok; SOCKET_PATCH_BUN_REQUIRED=1 cargo test -p socket-patch-cli --all-features --test e2e_vendor_bun_build 40/40 ok. cargo test --workspace --all-features --no-fail-fast: 241 binaries ok, 10 failures, none in code this PR touches. Nine are permission-denial tests (unwritable/unremovable files), which can't fail as intended because the sandbox runs as root. The tenth is e2e_bun_lockb::binary_shared_bundled_record_hosted_pin_is_managed, a hosted rollback that needs patches-api.socket.dev, which the sandbox proxy blocks. CI is the arbiter for these.

Review notes (/code-review high)

• Fixed: mirror_metadata read every sibling before checking its name; the doc comment overstated what can outrank the uuid file.
• Not changed: the LFS e2e soft-skips when git-lfs is missing even under the required gate. The Depot runner images aren't guaranteed to ship git-lfs, and without it the test can't reproduce the bug at all. The core check-attr test covers the mechanism on every runner.
• Not changed: an in-sync re-run that upgrades a legacy .gitattributes emits no event. It's a one-time metadata rewrite of a file that isn't part of the artifact, the same as today's restore of an edited file.

Follow-ups (not in this PR)

• JVM trees (vendor/jvm/mod.rs TREE_GITATTRIBUTES, redownload.rs) and the Gradle redirect still write * -text, so *.jar filter=lfs repos are likely affected the same way (p3, separate ownership checks).
• A tarball already committed as an LFS pointer isn't detected; a git cat-file pointer probe with a warning could replace the documented --renormalize step.
• .git/info/attributes, and a .gitattributes shipped inside a vlt directory payload, can still override the uuid file.

🤖 Generated with Claude Code

https://claude.ai/code/session_01WdaShef4vyGWjbTSpLbh9E

Generated by Claude Code