GeezerVison

Problem

Architecture audit §3.A, B31: which lockfile drives a project's installs was decided in at least eight places, each a hand copy. The copies had already drifted:

• the vendored npm router (vendor/npm_flavor.rs) and its hand-copied in-memory twin (lock_inventory/view.rs), which had lost the pnpm-PnP carve-out and the vendor_multiple_lockfiles warnings;
• the inventory's migration-leftover fallback (inventory_live_sibling_lock_in), which described itself as "a literal transcription of the router's order";
• hosted vlt's SIBLING_LOCKS list (redirect/vlt.rs), the hosted vlt preflight inputs (hosted/vlt.rs), and the hosted npm rewriter's "another lock owns it" hand list (redirect/mod.rs);
• the hosted governing-root check's own lock lists (hosted/governing_root.rs): OWN_LOCKS + NPM_LOCKS + VLT_LOCK for "the project is its own lock root", and WORKSPACE_ROOT_LOCKS for npm/yarn/Bun workspace roots;
• the PyPI order in the vendored router (detect_pypi_flavor), the hosted pdm_drives gate, and the agent-mode PDM crawler (pdm_drives_project).

The concrete user harm is #612. Vendored PyPI wires one lock. A requirements.txt exported beside the winning tool lock (pipenv requirements, uv export) was never counted as a losing install source, so:

• vendoring left it unpatched with no warning;
• a hosted→vendored takeover restored its hosted pin to upstream without a word.

Change

1. One table: formats/governing_locks.rs. It is pure and presence-only, so it passes the formats purity guard. It holds:

   • the npm-family precedence (NPM_PRECEDENCE: vlt > bun > pnpm > yarn > npm, as NpmLockFamily with each family's files);
   • the PyPI tool-lock precedence (PYPI_TOOL_LOCKS: uv > poetry > pdm > Pipfile);
   • small queries over both: npm_governing_family, npm_locks_outside, pypi_governing_tool_lock, pypi_tool_lock_governs, pypi_tool_lock_shadowed, pypi_locks_outside.

   Content refinements still belong to the router that reads the bytes: pnpm v9 vs legacy, yarn classic vs berry, the vlt version sniff, and pylock-contains-target.

2. npm family.

   • The vendored router is now written once, over ProjectView, as detect_npm_lock_flavor_in. The disk entry point is a thin wrapper around it, and the in-memory copy in view.rs is deleted.
   • The pnpm-PnP layout check is now pnpm_pnp_layout_in(view), which uses the new ProjectView::is_dir. The disk pnpm_pnp_layout wraps it.
   • inventory_live_sibling_lock_in, vlt::sibling_locks, the hosted vlt inputs and rewrite_npm_lock's sibling check all ask the table.
   • hosted/governing_root.rs derives its own-lock-root list from npm_lock_files() (plus pnpm's pre-v3 shrinkwrap.yaml, which the table does not carry) and its workspace-root list from the Npm, Yarn and Bun families' files().

3. PyPI. detect_pypi_flavor, pdm_drives and the PDM crawler's pdm_drives_project all ask the table.

4. Vendored mode in a Pipenv project wires only Pipfile.lock and silently leaves a sibling requirements.txt unpatched, and the hosted → vendored takeover reverts that file's hosted pin to plain PyPI #612 fix (behavior change, failing-first test). A root requirements.txt that names the package being vendored, sitting beside the governing tool lock, is now listed among the losers of the documented pypi_multiple_lockfiles warning: "wiring Pipfile.lock — installs driven by requirements.txt will still install the UNPATCHED registry bytes". The takeover path runs the same router after its restore, so a hosted→vendored takeover that unpatches requirements.txt now says so too; a CLI test (mode_migration_pypi::pipenv_hosted_to_vendored_names_the_unpatched_requirements) pins that path end to end. vendor --check is not changed by this PR. CLI_CONTRACT.md's row for the code now mentions it.

Duplicate copies deleted

Behavior changes (where copies disagreed or output order moved)

• Vendored mode in a Pipenv project wires only Pipfile.lock and silently leaves a sibling requirements.txt unpatched, and the hosted → vendored takeover reverts that file's hosted pin to plain PyPI #612: the new pypi_multiple_lockfiles warning described above.
• In-memory npm router. It now uses the same table as disk, which changes three things:
  • It emits vendor_multiple_lockfiles warnings. Its only caller, the inventory, discards them.
  • It applies the pnpm-PnP carve-out when a memory project carries an installed pnpm store. Both outcomes are refusals the inventory propagates.
  • Its Rush and missing-lock refusal details now match the disk wording.
• Warning order follows precedence order: vendor_multiple_lockfiles per file, and the file list in hosted redirect_vlt_sibling_lockfiles. The wording is unchanged. In redirect_workspace_lockfile_elsewhere, a workspace root holding both npm locks now lists npm-shrinkwrap.json before package-lock.json (npm's own preference); nothing else in that message moves.
• Snapshot views now read the sniffed lock through the snapshot cache instead of an uncached disk read. Within one run the bytes are the same.

Testing

All runs on macOS, after a rebase onto origin/main 6fe81ad, using heavy-job.sh with CARGO_INCREMENTAL=0 and -j4.

• Failing first (unit). cargo test -p socket-patch-core --lib -- requirements_beside_the_governing_lock failed before the Vendored mode in a Pipenv project wires only Pipfile.lock and silently leaves a sibling requirements.txt unpatched, and the hosted → vendored takeover reverts that file's hosted pin to plain PyPI #612 change (Pipfile.lock: [], 0 warnings) and passes after it.
• Failing first (CLI takeover). cargo test -p socket-patch-cli --test mode_migration_pypi pipenv_hosted_to_vendored_names_the_unpatched_requirements passes, and fails at the warning assertion with the Vendored mode in a Pipenv project wires only Pipfile.lock and silently leaves a sibling requirements.txt unpatched, and the hosted → vendored takeover reverts that file's hosted pin to plain PyPI #612 requirements probe disabled (checked by a temporary local mutation, reverted).
• Core lib. cargo test -p socket-patch-core --lib: 5582 passed, 0 failed. This includes formats::governing_locks, the formats purity guard, hosted::governing_root, the npm-flavor tests, the in-memory flavor test (now covering the pnpm-PnP carve-out for both store markers, and yarn.lock overriding it), and the inventory, vlt, pdm and pypi suites.
• Core integration. cargo test -p socket-patch-core --test redirect_golden --test hosted_inventory: all pass.
• CLI. cargo test -p socket-patch-cli --test mode_migration_pypi (32/32) --test in_process_vendor_pypi_takeover (6/6) --test in_process_redirect (117/117) --test in_process_redirect_pipenv (9/9) --test in_process_redirect_pnpm (19/19, covers the workspace-root refusals); earlier rounds also ran mode_migration_vlt, in_process_vendor, e2e_vendor_pypi_build, e2e_vendored_production: all pass.
• mode_migration_npm. 18 of 19 pass. berry_vendored_then_hosted_takeover_leaves_pure_hosted fails identically from a build without this change, so the failure is not from this PR. It only runs where corepack can fetch yarn berry: hosted berry pins through resolutions, but the test expects a pristine package.json.
• Clippy. cargo clippy -p socket-patch-core -p socket-patch-cli --all-targets -- -D warnings reports nothing on the lines this PR touches. Main's existing hits (python_crawler, jvm_jar, maven_repo, nuget_feed and others) remain.
• Format. rustfmt --check is clean on every touched file.
• Left to CI. The Linux and docker suites.

Deferred (policy, not decided here)

• vlt beside another lock. With vlt-lock.json and pnpm-lock.yaml present and no node_modules/.vlt-lock.json sentinel, vendored wires vlt, while hosted's vlt_drives lets neither lock decide alone. Both rules now read the same family table, but choosing one rule is a maintainer decision.
• Hosted Pipenv veto. pipenv::rewrite still uses its own liveness rule (a Pipfile beside it) rather than the tool-lock precedence. A leftover Pipfile.lock beside uv.lock can still veto. Aligning it with pypi_tool_lock_governs would change hosted refusals.
• pdm_drives and pylock. pdm_drives still ignores standalone pylock, which vendored ranks above poetry when it pins the target. Adding it would make hosted rewrite fewer files.
• Takeover restore scope. A hosted→vendored takeover still restores every hosted file, including the losers (Vendored mode in a Pipenv project wires only Pipfile.lock and silently leaves a sibling requirements.txt unpatched, and the hosted → vendored takeover reverts that file's hosted pin to plain PyPI #612's "never leave a file less patched"). Keeping a hosted pin in an unwired file means one purl is in two modes at once. That needs a decision on mixed-mode ledgers (related to C34). This PR makes the outcome loud instead of silent.
• Install-state detection. crawlers/pkg_managers::detect_npm_pkg_manager (agent mode) answers a different question, which package manager installed node_modules, so it keeps its own order. It shares only the PnP predicate.
• Bun presence probes (E17, Lock inventory ignores a bun.lockb that Bun installs from when bun.lock is a dangling symlink #735) are left to the bun session (Fix open bun issues (#992, #861, #784, #764, #735, #635, #599, #578, #497, #443, #371) #1009). The table keeps today's exists semantics.
• requirements.txt includes and non-pin lines. The Vendored mode in a Pipenv project wires only Pipfile.lock and silently leaves a sibling requirements.txt unpatched, and the hosted → vendored takeover reverts that file's hosted pin to plain PyPI #612 warning reads only the root requirements.txt, and only its named-requirement lines: -r/--requirement includes, -e editable installs, and bare path or URL lines are not probed, so the target installed through one of those stays silent.

Related: #612 (partly addressed: the silent case is now loud; the takeover restore scope is deferred, so this does not close it).

🤖 Generated with Claude Code

Generated by Claude Code
d CRLF/LF outside edited spans and treating $ in URLs literally. Berry writers use BerryStanzas for BOM, EOL, and sorted re-insertion.

B16 behavior change: CopySource distinguishes registry copies from remote tarballs (file: tgz, URL forks, codeload shorthands). Non-registry blocks are no longer repointed to Socket’s registry artifact; they are skipped with new warnings, excluded from VEX where applicable, and rollback refuses legacy hosted pins on those keys. Lock inventory drops registry verifiers for fork copies. Docs in docs/ecosystems.md are updated.

^{Reviewed by Cursor Bugbot for commit 209e36e. Configure here.}

Generated by Claude Code