Vendored scan of a fresh uv checkout (uv.lock, no .venv yet) exits 1 on packages that exist only in the system Python, because the crawler falls back to the global site-packages (the uv side of #947) #964

[agent] Found by the scheduled uv bug-hunt routine (ledger #310).

Summary

This is the uv counterpart of #947 (Pipenv), and the open fix PR for #947 (#950) doesn't cover it. #950 only stops the fallback for Pipenv projects (is_pipenv_project → empty). It keeps crawler_python_e2e.rs::get_site_packages_paths_falls_back_via_uv_lock_marker, which pins the uv.lock → global fallback.

In a uv project checked out fresh (pyproject.toml + uv.lock, before uv sync, so there's no .venv), PythonCrawler::get_site_packages_paths finds no venv. is_python_project(cwd) is then true because of uv.lock / pyproject.toml, so the crawl falls back to get_global_python_site_packages(). Every package in the OS Python joins the candidate set (scannedPackages: 144 for a 2-package lock). scan --mode vendored then tries to vendor the system-only ones into uv.lock and fails:

pkg:pypi/pyyaml@6.0.1  failed  pypi_uv_lock_package_missing
  "uv.lock has no [[package]] entry for pyyaml; run `uv lock` first"
status: partial_failure, exit 1

The suggested remedy can't help, because pyyaml isn't a project dependency. --dry-run exits 0 with status: success and doesn't flag the refusal.

Impact

• CI breakage in the documented vendored flow. A fresh checkout is exactly where vendored mode runs. GitHub ubuntu-* runners and Debian/Ubuntu images ship a system Python with PyYAML, requests, urllib3, jinja2, cryptography, six, idna and more. Whenever Socket has a patch for any of them, every vendored scan of every uv project on that machine exits 1. That includes projects whose own dependencies have no patches at all (repro below: nothing in the lock is patched, and the run still fails).
• A system package that shares a name with a lock package masks the lock version. Here the system six 1.16.0 counts as "installed", so lockfileOnlyPackages is 1 instead of 2, and the system copy's bytes stand in for the project's.
• Patch counts, the rollout budget (--max-new-patches) and the report include packages that aren't part of the project.
• Nothing outside the project is written in vendored or hosted mode. The agent-mode write to the system Python is the Agent-mode scan in a Pipenv project without a Pipenv venv patches the system Python's site-packages in place instead of the project's venv/ (regression from #388) #504 shape, which the uv ledger tracks as a known generic fallback.

Repro (Linux, Debian/Ubuntu system Python with python3-yaml 6.0.1; a mock public proxy, --proxy-url, offers one free patch for pkg:pypi/pyyaml@6.0.1 and nothing else)

mkdir app && cd app
cat > pyproject.toml <<'EOF'
[project]
name = "app"
version = "0.1.0"
requires-python = ">=3.9"
dependencies = ["six==1.16.0", "idna==3.7"]
EOF
uv lock                      # lock only: a fresh checkout, no .venv
env -u VIRTUAL_ENV socket-patch scan --mode vendored --yes --ecosystems pypi --json \
    --proxy-url http://127.0.0.1:8767 > out.json; echo "exit=$?"
# exit=1, status partial_failure, scannedPackages 144, lockfileOnlyPackages 1
# vendor.events: pkg:pypi/pyyaml@6.0.1 failed pypi_uv_lock_package_missing
python3 -c "import yaml; print(yaml.__file__)"   # /usr/lib/python3/dist-packages/yaml/__init__.py

Control: run uv sync first, so .venv exists. The same scan crawls only the venv (scannedPackages: 2, lockfileOnlyPackages: 0), finds no patch, and exits 0. With a mock that also patches six, the run fails the same way for pyyaml while six goes down the normal vendoring path.

Expected vs actual

• Expected: CLI_CONTRACT.md, "cwd-only (single project)", says the pypi crawler inspects only the project rooted at --cwd: the manager-recorded env (UV_PROJECT_ENVIRONMENT), $VIRTUAL_ENV, .venv / venv. Global packages are the -g / --global-prefix surface. Lock-only packages already come from uv.lock (lockfileOnlyPackages), so on a fresh checkout the lock is the whole candidate set. The run should vendor what the lock holds, exit 0, and never mention pyyaml.
• Actual: system-Python packages join the candidate set. Vendored exits 1 with pypi_uv_lock_package_missing and a misleading uv lock remedy, --dry-run reports success, and system copies mask lock entries.

Cells (main 9c43dfc, CLI 4.0.0, Linux x86_64, system Python 3.11/3.12 dist-packages)

The behaviour doesn't depend on the uv version: uv only writes the lock, and the crawl is the same for every lock grammar. macOS and Windows weren't probed (the fallback is cross-platform: get_global_python_site_packages also scans Homebrew, python.org and conda roots). Not bisected; the fallback predates vendored uv support.

Suspect code

• crates/socket-patch-core/src/crawlers/python_crawler.rs:3047: if is_python_project(&options.cwd).await { return Ok(get_global_python_site_packages().await); } in non-global mode. The doc comment above it ("a fresh clone with pyproject.toml + uv.lock but no .venv would silently return zero packages") predates lock-only discovery. That case is now covered by the lock merge.
• crates/socket-patch-core/tests/crawler_python_e2e.rs:1124: get_site_packages_paths_falls_back_via_uv_lock_marker pins the fallback for uv.lock. PR Fix Pipenv project falling back to system Python (#504, #947) #950 leaves it in place.
• crates/socket-patch-core/src/vendor/pypi_uv.rs:330-334: the pypi_uv_lock_package_missing detail recommends uv lock, even when the package isn't a project dependency.

Related: #947 / PR #950 (the Pipenv half; this report came from the Pipenv routine's handover), #504 (agent-mode writes), #476 (Poetry), #502 (PDM).