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

Summary

After vendoring a PEP 723 script (tool.py + tool.py.lock), adding any unrelated dependency with uv add --script tool.py <pkg> makes vendor --revert and rollback keep the vendored wiring for good. They report vendor_lock_entry_drifted, but the patched package's own lock entry didn't change. uv only added one line to the script lock's [manifest] requirements array.

The revert's three-way merge in vendor/pypi_lock.rs treats any length change in a recorded array as drift. The script lock's [manifest] requirements is recorded as a whole array (it holds six's path entry), so a new requirement from the user counts as a conflicting edit. The same flow on a project uv.lock (uv add idna after vendoring) reverts cleanly.

Impact

• vendor --revert exits 0 with status: success, but the script and lock still point at .socket/vendor/…, and .socket/vendor/pypi/<uuid> and the ledger entry are kept. uv run --locked --script tool.py keeps running the patched wheel.
• rollback exits 1 (partial_failure, vendoredKept: lockfile wiring drifted).
• The advice "undo the drift (restore the vendored lock entries or re-vendor) and re-run vendor --revert" can only be followed by dropping the user's new dependency. Re-vendoring doesn't help: scan --mode vendored reports already_vendored, and the next vendor --revert drifts again (tested).

Repro (Linux, main 6e7ef74)

Same local mock patch API as the other uv issues (embedded in the probe workflow below).

cat > tool.py <<'PY'
# /// script
# requires-python = ">=3.9"
# dependencies = ["six==1.16.0", "python-dateutil==2.8.2"]
# ///
import six; print("PATCHED" if getattr(six, "SOCKET_PATCHED", 0) else "PRISTINE")
PY
uv lock --script tool.py && git init -q && git add -A && git commit -qm init
socket-patch scan --mode vendored --json --yes $API   # applied; uv run --locked --script tool.py → PATCHED
uv add --script tool.py idna==3.7                      # only adds idna to [manifest] requirements + a [[package]]
socket-patch vendor --revert --json --yes $API         # exit 0, success, events: vendor_lock_entry_drifted, vendor_artifact_kept, vendor_revert_kept
grep -c socket/vendor tool.py tool.py.lock             # 1 / 2: still vendored
uv run --locked --script tool.py                       # PATCHED

Isolation, starting from the vendored state, with one edit each:

Expected vs actual

• Expected: CLI_CONTRACT.md: "vendor --revert restores the originals (fragments that no longer match — a user re-resolved — are left alone with a vendor_lock_entry_drifted warning …)". Nobody re-resolved six here, and its [[package]] entry and [manifest] element are exactly what vendoring wrote. Only a sibling element was added, so the revert should put six's path element back to its specifier form and leave idna alone, as the project-lock backend does.
• Actual: the whole revert is skipped and the artifact is kept. vendor --revert still exits 0 with success.

OS × uv matrix

First bad

Not bisected. The only published release (4.0.0) predates the v5 vendoring consolidation (#277), so it isn't a usable comparison point.

Suspect code

crates/socket-patch-core/src/vendor/pypi_lock.rs:478 (restore_value: if live.len() != new.len() || original.len() != new.len() { return true; }), reached through restore_document for the lock file in revert_python_locks (pypi_lock.rs:600-615). Array elements should be matched by identity (name, plus marker/extras where needed), not by position and length. Line 513 has the same length check for arrays of tables.

Probe: https://github.com/SocketDev/socket-patch/actions/runs/36878632878 (ubuntu, macos, windows × uv 0.5.31 / 0.12.21)