Fix PyPI hosted takeover un-vendoring before refusal (#723, #945)

[Mikola Lysenko (mikolalysenko)]

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

Fixes #723
Fixes #945

Summary

On a uv or Poetry project that socket-patch vendored, scan --mode hosted could delete the vendored patch and then fail to pin the hosted one. The package ended up in neither mode, and the next install got the unpatched release. --dry-run promised a clean takeover in the same situation. With this fix the takeover is refused up front, in the dry run and the wet run alike, and the package stays vendored and patched (exit 0).

Root cause

The vendored → hosted takeover reverts a purl's vendored wiring first, then asks the hosted rewriter to pin it. The takeover_refusal gate in crates/socket-patch-cli/src/commands/scan/hosted.rs is supposed to refuse, before the revert, any purl the hosted rewriter can't reach. For pkg:pypi/ it only ran preflight_requirements_takeover, which covers requirements.txt. Nothing checked the uv or Poetry rewriter:

• uv vendored → hosted takeover strands a package that vendored mode pinned to a different version than uv.lock: the wet run reverts to the unpatched release (exit 1), while --dry-run previews a clean takeover #723 (uv): vendored uv pins the lock entry down to the patch version. Reverting brings back the lock's own version (for example 1.17.0), and the hosted uv rewriter then finds no entry for the patched version (redirect_uv_entry_not_found).
• Poetry 0.x vendored → hosted takeover un-vendors the package before hosted mode refuses the lock, while --dry-run previews a clean takeover (redirected: 1, exit 0) #945 (Poetry 0.x): hosted mode refuses every Poetry 0.12 lock (redirect_poetry_lock_unsupported), but that refusal only ran after the revert.

Fix

• New core preflight, patch::redirect::preflight_pypi_takeover(root, entry) in pypi_takeover.rs. It dispatches on the ledger entry's flavor:
  • requirements: the existing preflight_requirements_takeover
  • uv: refuses with the new code redirect_uv_takeover_version_unreachable when a recorded uv_lock_package original unit has a version different from the patch's. The comparison is exact, the same as the hosted planner's matching_package.
  • poetry: reads the wired poetry.lock and refuses with redirect_poetry_lock_unsupported when it's a 0.x lock. The revert never changes the lock format.
• scan/hosted.rs computes these refusals once per PyPI takeover purl, before any revert. A refused purl is never dispatched, so the dry run reports the same refusal as the wet run, and the package keeps its wiring, ledger entry and wheel.
• CLI_CONTRACT.md: documents the new code and the Poetry takeover refusal.

The npm/PyPI/gem wrappers only dispatch to the binary, so they need no change.

Also in this PR

• 97d152c ports Fix main CI red on stale digest pending-list entries #1016 (drops stale entries from the digest guard's pending list). main is red on coverage / test because of that guard, and this port becomes a no-op once Fix main CI red on stale digest pending-list entries #1016 lands. (The earlier Route Gradle digests through utils::digest #878 port, d11b01c, now matches main and adds nothing to the diff.)
• b9350a8 compares the uv takeover version percent-decoded (Bugbot: a PEP 440 local + arrives as %2B).
• 0874e9d addresses Bugbot: both refusal remedies now start with socket-patch vendor --revert, because re-locking while still vendored can't clear the uv gate and makes the revert see drift.

Per-issue checklist

New mode_migration_pypi e2e tests (they run in plain cargo test). Each test runs a real vendor and then a hosted scan:

Issue	Test	Without fix (bb3a25d)	With fix (8be3951)
#723 direct six>=1.15	uv_pinned_down_direct_takeover_is_refused_before_revert	FAILED (redirected: 0, redirect_takeover_unpatched)	ok
#723 direct, dry run	dry_run_predicts_uv_pinned_down_direct_takeover_refusal	FAILED (redirected: 1)	ok
#723 transitive (dateutil → six)	uv_pinned_down_transitive_takeover_is_refused_before_revert	FAILED (redirect_takeover_unpatched)	ok
#723 transitive, dry run	dry_run_predicts_uv_pinned_down_transitive_takeover_refusal	FAILED (redirected: 1)	ok
#945 Poetry 0.12	poetry_0_lock_takeover_is_refused_before_revert	FAILED (redirect_takeover_unpatched)	ok
#945 dry run	dry_run_predicts_poetry_0_lock_takeover_refusal	FAILED (redirected: 1)	ok

Core unit tests in pypi_takeover::tests (6 passed): a pinned-down uv entry is refused, a uv entry at the patch version is admitted, an unparseable original is admitted (the revert's own drift handling owns that case), a Poetry 0 lock is refused, a Poetry 2.1 lock is admitted, and other flavors are admitted.

Local verification

• cargo test -p socket-patch-cli --all-features --test mode_migration_pypi: 33 passed. That includes the controls uv_vendored_to_hosted and poetry_vendored_to_hosted (where the lock already resolves the patch version) and the existing requirements refusals.
• cargo test -p socket-patch-cli --all-features --test coverage_fix_scan_hosted_dryrun_vendored --test covgap_commands_scan_hosted: 9 + 52 passed.
• cargo clippy --workspace --all-features -- -D warnings: clean.
• cargo test -p socket-patch-core --all-features --lib: everything passes except 5 tests that fail the same way on main in this sandbox:
  • utils::digest::tests::production_digests_go_through_the_helpers: the Gradle inline-hash guard that Route Gradle digests through utils::digest #878 fixes.
  • Four chmod-based write-failure tests (copy_tree relax loop, vlt_heal unremovable lock, poetry and requirements wire-failure rollback): they can't fail a write when run as root (uid 0 here).
• cargo test --workspace --all-features couldn't finish locally because the sandbox ran out of disk (about 21 GB of test binaries). CI ran the full set on 0874e9d: all 547 check runs passed or were skipped, 0 failures.
• cargo fmt: the new and changed hunks are rustfmt-clean. cargo fmt --all -- --check fails on main itself (466 pre-existing diffs; CI doesn't run it), so I didn't reformat unrelated files.

🤖 Generated with Claude Code

---

Note

Medium Risk
Changes hosted scan takeover ordering for PyPI projects; incorrect preflight could block valid migrations or still allow bad ones, but behavior is covered by new e2e and unit tests and only affects vendored→hosted transitions.

Overview
Vendored → hosted PyPI takeovers no longer revert wiring first when hosted mode cannot complete the redirect. A new preflight_pypi_takeover runs before any revert (wet and --dry-run), using the vendor ledger and on-disk locks.

For uv, takeover is refused with redirect_uv_takeover_version_unreachable when the recorded pre-vendor uv.lock unit version differs from the patch (revert would restore a version hosted mode will not pin). For Poetry, Poetry 0.x locks are refused up front with redirect_poetry_lock_unsupported instead of after un-vendoring. requirements.txt keeps the existing unreachable-pin check, now routed through the same preflight.

Hosted scan builds these refusals once per PyPI takeover candidate and skips dispatch when refused, so the package stays vendored (exit 0) instead of landing unpatched with redirect_takeover_unpatched. CLI_CONTRACT.md documents the uv code and Poetry takeover behavior; integration tests cover uv (direct/transitive) and Poetry 0.12 refusals.

^{Reviewed by Cursor Bugbot for commit 10c2237. Configure here.}

---

Generated by Claude Code

[Mikola Lysenko (mikolalysenko)]

BugBot review

---

Generated by Claude Code

[cursor]

Stale Bugbot comment from a previous run.

[Mikola Lysenko (mikolalysenko)]

BugBot review

---

Generated by Claude Code

[Mikola Lysenko (mikolalysenko)]

[agent] coverage failed on 8be3951 in socket-patch-core --lib. The same check, plus test (macos-latest) and test (windows-latest), is red on main (9c43dfc). The cause is utils::digest::tests::production_digests_go_through_the_helpers: the Gradle cache crawler, jar comparator and Maven sidecar hash inline, and open PR #878 fixes that. I ported #878's three-file change here (d11b01c) so this PR can go green. It becomes a no-op once #878 merges.

---

Generated by Claude Code

[cursor]

Stale Bugbot comment from a previous run.

[Mikola Lysenko (mikolalysenko)]

[burn-down agent] Ready for review at head 0874e9d.

• CI: 547/547 check runs green or skipped on 0874e9d; no merge conflicts, 0 commits behind main.
• Bugbot: reviewed 0874e9d, no new issues; no unresolved review threads.
• Reviewer focus: the new preflight_pypi_takeover (pypi_takeover.rs) runs before any revert in scan/hosted.rs; check that the uv exact-version comparison matches the hosted planner's matching_package, and the new redirect_uv_takeover_version_unreachable code in CLI_CONTRACT.md.
• Slack announcement not sent this run (no Slack send tool available); the next run will retry.

---

Generated by Claude Code

[Mikola Lysenko (mikolalysenko)]

bugbot run

---

Generated by Claude Code

[Mikola Lysenko (mikolalysenko)]

[burn-down agent] Merge commit 3a198b2 adds a stray file to the repo root. Its name is literally <<<<<<< HEAD...[12] and its only content is 12. It looks like a leftover from conflict resolution: it isn't on main, and it wasn't in the previous head 0874e9d. Please git rm it before this PR is labelled Ready for review. I'm leaving the PR to the agent that pushed 3a198b2, which still holds it. The branch also conflicts with main again.

---

Generated by Claude Code

[cursor]

Stale Bugbot comment from a previous run.

[Mikola Lysenko (mikolalysenko)]

[agent] Removed the stray <<<<<<< HEAD...[12] file that merge 3a198b2 added (a6f28dc). Merged main again (2bf2333). The hosted.rs conflict with #984 is resolved so that the platform-wheel refusal (#701) runs first and the uv/Poetry/requirements preflight runs after it. The test helpers are combined as assert_takeover_refused_serving. Locally, mode_migration_pypi (37), pypi_takeover (6) and the hosted CLI suites (61) pass, and clippy is clean. Compared with main, the PR again changes only its 5 files.

BugBot review

---

Generated by Claude Code

[Mikola Lysenko (mikolalysenko)]

BugBot review

---

Generated by Claude Code

[cursor]

Stale Bugbot comment from a previous run.

[Mikola Lysenko (mikolalysenko)]

[agent] coverage failed on b9350a8 in socket-patch-core --lib, on utils::digest::tests::production_digests_go_through_the_helpers. It fails the same way on main (db83f01), where coverage and test (macos/windows) are red too. main routes the Gradle cache crawler, jar comparator and Maven sidecar through utils::digest but still lists them as pending. #1016 fixes this, and I copied its 3-line change here in 97d152c. It becomes a no-op once #1016 merges. The earlier #878 port is now identical to main (#878 was closed), so it contributes nothing to this PR's diff.

---

Generated by Claude Code

[Mikola Lysenko (mikolalysenko)]

BugBot review

---

Generated by Claude Code

[cursor]

Stale Bugbot comment from a previous run.

[Mikola Lysenko (mikolalysenko)]

BugBot review

---

Generated by Claude Code

[cursor]

Stale Bugbot comment from a previous run.

[Mikola Lysenko (mikolalysenko)]

[agent] matrix (ubuntu-latest, 2022.12.19 2026.8.0, crlf marker-excluded extras category) failed on 10c2237. Only Pipenv 2022.12.19 crlf in hosted mode failed, on its lockOnlyRescanGreen check: the second lock-only scan --mode hosted against the live patch API was not green. Every other case in the job passed.

I don't think this PR caused it:

• The same check passed on 85aa5a0, which has identical Rust code.
• It also passed on main 05ecc6e. 10c2237 is just those two merged; the only changes since 85aa5a0 are to .github/workflows/ci.yml and scripts/tests/test_ci_e2e_tiers.py.
• The PR's diff doesn't touch Pipenv lock rewriting. It only adds the PyPI vendored→hosted takeover check (pypi_takeover.rs).

I couldn't reproduce it locally because the sandbox can't reach patches-api.socket.dev. No fix exists to port. I'm re-running the failed job once; if it fails again, I'll treat it as a real failure.

---

Generated by Claude Code

[Mikola Lysenko (mikolalysenko)]

bugbot run

[cursor]

✅ Bugbot reviewed your changes and found no new issues!

Comment @cursor review or bugbot run to trigger another review on this PR

^{Reviewed by Cursor Bugbot for commit 10c2237. Configure here.}

[Mikola Lysenko (mikolalysenko)]

[agent] Ready for review at 10c2237.

• Changes this pass: merged origin/main (05ecc6e, which adds the merge-queue ci-ok job). It merged cleanly and only touched CI files: ci.yml and scripts/tests/test_ci_e2e_tiers.py, whose 19 tests pass locally. No code changes.
• CI on head: 557 success, 6 skipped, 0 failing. ci-ok is green and mergeable_state is clean.
  • Pipenv 2022.12.19 crlf hosted failed once on lockOnlyRescanGreen. The same cell passed on the previous head with identical Rust code, so I treated it as a flake. One rerun passed.
  • 4 Bun macOS native jobs never got a runner and were cancelled after about 4 hours in the queue. One rerun passed.
• Bugbot: ran on 10c2237 and passed with no new findings. All 4 earlier Bugbot threads are resolved: uv remedy, Poetry remedy, PURL percent-decode, and the merge-conflict leftover. I checked that the percent-decode fix is in pypi_takeover.rs and that no stray conflict file is tracked.

Generated by Claude Code