GeezerVison

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

Fixes #832
Fixes #958

Summary

The vendored-reference scan (scan_vendor_references) is what keeps a .socket/vendor/<eco>/<uuid>/ unit with no ledger entry from being deleted while a project file still points at it. Every caller goes through it: repair (vendor_ledger_missing), the orphan sweeps (vendor --revert, the vendored gc), the vendor stranded-reference gate, and rollback's ledger-less gate. It missed NuGet, Maven and Hatch wiring for two reasons:

1. Two notions of "a file a vendored run writes". The scan read registry::paths_with(VENDORED). nuget.config (all three spellings), packages.lock.json, pom.xml, .mvn/maven.config, hatch.toml and pnpm-workspace.yaml sat in a second list, VENDORED_WRITES_UNMARKED, that only the vendored dry run's symlink check (wiring_paths) read.
2. The grammar. NuGet's feed (value=".socket/vendor/nuget/<uuid>") and Maven's repository (<url>file://${project.basedir}/.socket/vendor/maven/<uuid></url>) name the uuid directory itself, while parse_vendor_path requires a leaf. The scanner also had no < terminator.

Why (leverage)

B 2 (#832, #958: same root cause, per the fixer's "Shares root cause" comments), U 0, D ≈1 (two registry notions collapse to one role, and the second list is deleted), R M. Score ≈5, the best eligible candidate this run: every higher-ranked queue row is blocked by open PRs on commands/vendor.rs or by claims. Register rows E61 and E67 (register); living document doc/05-vendored.md, new findings E61/E67.

What changed

• formats/registry.rs: the eight rows get the VENDORED role, and VENDORED_WRITES_UNMARKED is deleted. wiring_paths(eco) is now just ecosystem && VENDORED, and its output is unchanged for every ecosystem.
• vendor/path.rs: new parse_vendor_reference, which is parse_vendor_path's grammar but also accepts an empty leaf (the bare uuid dir). parse_vendor_path is now parse_vendor_reference(..).filter(leaf non-empty), so there is one grammar and every existing caller is unchanged.
• vendored_backend/repair.rs: the scan reads through parse_vendor_reference, and < ends a reference. A directory-wired unit reports the uuid dir as its path, with any trailing / trimmed.
• utils/digest.rs (separate commit): ports the base-red fix that Honor HTTP-date Retry-After on vendor-service retries through api::retry (#677) #889 and Fix uv dry run missing inline [tool.uv] refusal (#979) #980 carry. main fails production_digests_go_through_the_helpers because sbt, Mill and scala-cli support in agent, hosted and vendored modes #690 migrated gradle_cache.rs, jvm_jar.rs and sidecars/maven.rs while the pending list still names them. It is the same three-line removal, and it no-ops once main carries it.

Deleted

git diff --stat origin/main, split at #[cfg(test)]:

• production: +41 / −41 (VENDORED_WRITES_UNMARKED and its filter are gone; the grammar is shared)
• tests: +255 / −0 (plus −3 in the digest ratchet list)

Behavior

These are the intended fixes:

• With a missing ledger entry, a unit wired by nuget.config, pom.xml, hatch.toml, packages.lock.json, .mvn/maven.config or pnpm-workspace.yaml is now kept by the orphan sweeps (stillWired) instead of being deleted.
• repair now reports such a unit as vendor_ledger_missing (exit 1) instead of finding no vendored traces.
• The vendor stranded-reference gate and rollback's ledger-less refusal now see the same references.

Nothing else changes: wiring_paths returns the same files, and parse_vendor_path callers see the same results. Not done here: the dead eco == "maven2" arm of the stranded-reference gate in commands/vendor.rs, which is unreachable. That file is changed by #776 and #978, so the arm is left for a follow-up.

Test evidence

• New tests:
  • repair_vendor_e2e::repair_reports_missing_ledger_for_nuget_maven_and_hatch_wiring (CLI, built binary): on main it fails with left: [] (no events); on this branch it passes with 3 vendor_ledger_missing events (maven, nuget, pypi).
  • vendored_backend::repair::tests::scan_recovers_unit_dir_and_hatch_toml_references: covers the NuGet dir with a trailing /, the backslashed NuGet.Config, Maven <url>…</url> and hatch.toml. It also checks that a bare maven/, the v5 maven2 tree and a non-uuid dir are still not references.
  • vendor::path::tests::parse_vendor_reference_accepts_the_uuid_dir_itself: parse_vendor_path still rejects the bare dir.
  • formats::registry::tests::the_reference_scan_and_the_symlink_check_read_the_same_files: paths_with(VENDORED) equals the union of wiring_paths over every ecosystem, and the NuGet and PyPI lists are pinned.
• cargo clippy --workspace --all-features -- -D warnings: clean.
• cargo test -p socket-patch-core --lib: 5,559 passed. The 4 failures are the known root-only ones (copy_tree::relax_loop…, vlt_heal::an_unremovable…, pypi_poetry::wire_write_failure…, pypi_requirements::wire_failure…). The digest ratchet passes after the port.
• cargo test -p socket-patch-cli --all-features --test <suite>:
  • passing: repair (125 + 1 new), in_process_vendor (120), in_process_remove_repair_lifecycle (23), remove (91), rollback (38), scan (118), vendor_crash_safety_e2e (6), vendor_jvm_cli (29), e2e_sbt_vendor (18), mode_migration_npm (19), mode_migration_cargo (9), apply (106), cli (92);
  • root-only failures in the sandbox: repair's 2 known chmod tests, and covgap_commands_vendor's 3 *_state_write_failure_* tests, which chmod 0o555 the vendor dir and are ineffective as root (51 others pass).

Risk

M. Where a project file references a unit, the orphan sweep now keeps it rather than deleting it, so the risk is that something is kept, not that something is lost. The new < terminator and the bare-dir form only add matches, and a match still needs a known ecosystem dir and a canonical uuid.

🤖 Generated with Claude Code

https://claude.ai/code/session_01RcjbNmuocdNawny89vudHV