Harden uv lockfile patch preservation and verify every uv release family

[Mikola Lysenko (mikolalysenko)]

Follow-up to #238 (uv / PEP 751 / PEP 723 Python lock patch preservation). Fixes the findings from its review, fixes the uv-version-specific defects the extended matrix exposed, and re-runs the real-uv compatibility matrix against every 0.x release family — first release, latest release, and the releases on either side of every behaviour boundary — for both hosted and vendored modes, now including a project-variant lane (dev groups, duplicate requires-dist entries, constraints, transitive overrides) on every [[package]] binary.

Fixes from the #238 review

• vendor --revert no longer deletes a vendored Python wheel a lock still references. A ledger entry with no wiring records (the shape socket-patch repair re-synthesizes when state.json is lost) was routed into the uv / python-lock revert, iterated zero records, reported success, and the caller then removed the artifact directory and dropped the entry while uv.lock / pylock.toml / the script / requirements.txt still resolved through it. revert_pypi_opts now refuses with vendor_wiring_unknown_revert_blocked while any Python project file references the uuid directory, mirroring the npm-family guard.
• Inventory: a PEP 723 script lock or PEP 751 lock supplements the project's poetry.lock / requirements.txt pins instead of hiding them. uv.lock keeps its exclusive precedence.
• Discovery: symlinked locks are discovered (DirEntry::file_type does not follow symlinks); one unreadable directory entry no longer drops every lock.
• Line endings: CRLF pylock.toml, script locks and pyproject.toml keep their convention through rewrites and document-restore reverts.
• Ledger recovery compares the purl name in PEP 503 form; hosted pyproject.toml edits render uv's own [tool.uv.sources] layout; redirect_uv_missing_sha256 fires once per dependency; repair prefers uv.lock over an exported pylock when stamping flavor.

uv-version defects found by the extended matrix

The [[distribution]] lock grammar (uv 0.1.45–0.2.34) went through three shapes, and the hosted rewriter keyed everything on the table name:

uv	sources	artifacts
≤ 0.2.5	source = "registry+…" strings	[distribution.sdist] / [[distribution.wheel]] sub-tables
0.2.6 – 0.2.17	strings	inline sdist = {…} / wheels = [ … ]
0.2.18 – 0.2.34	inline tables { registry = … }	inline

The rewriter now decides the source shape and the artifact shape independently from the entry's own syntax; uv 0.2.35/0.2.36 locks without a root [package.metadata] vendor natively instead of being refused.

Second review pass (ultracode: 10 review dimensions, 3-lens adversarial verification, 4 real-binary probes)

Twenty-two findings survived verification; all are fixed here, each with a RED-first regression test. The two Cursor Bugbot findings are among them.

Correctness (uv end to end)

• [tool.uv] dev-dependencies is a direct dependency. It was classified transitive, wired through override-dependencies, and its [package.metadata.requires-dev] entry left stale, so uv sync --locked / uv lock --check failed on every uv ≥ 0.2.37 and a plain uv sync on 0.2.37/0.4.30 reinstalled the pristine wheel. Both the vendored classifier and the hosted rewrite_project_metadata now count it.
• Every duplicate requires-dist / requires-dev entry is repointed (dependencies + an extra, markers), not only the first — --locked accepted the lock only by accident before.
• [tool.uv] constraint-dependencies / build-constraint-dependencies: the lock's [manifest] constraints / build-constraints entries are repointed (hosted and vendored), the shape uv ≥ 0.5.6 writes; a new wiring kind with a revert arm. uv 0.2.37–0.5.3 reject the repointed entry under --locked, so the repoint emits the advisory pypi_uv_constraints_require_uv_0_5_6.
• Transitive (override) wiring advisory pypi_uv_override_requires_uv_0_5_6: uv applies [tool.uv.sources] to override-dependencies only from 0.5.6; on 0.2.35–0.5.3 --frozen installs the patch but a plain uv sync reinstalls the registry wheel. No lock marker separates those releases, so this is an advisory, not a refusal.
• Vendored uv backend is CRLF-preserving (pyproject, the rewritten [[package]] unit, appended [manifest] / [package.metadata] fragments, and their revert); previously a CRLF pair came back LF / mixed.
• [[distribution]] refusal names the real reasons (probed with real binaries: through 0.2.6 a relative path source does not parse; on 0.2.17–0.2.34 --locked rejects it and uv lock / plain uv sync absolutize it) instead of "records absolute paths". The [package.metadata] no-op gate also considers [package.metadata.*] sub-tables.

Symlinks (Bugbot finding 2 and its consequence)

• pypi_lock::read_file accepted only regular files by lstat, so a symlinked pylock.toml / *.py.lock — and, since discovery started following links, a stray non-referencing sibling symlink — refused vendoring of the whole project. It now reads through the link like every other reader.
• Every writer (hosted scan, vendored uv, python-lock) stages next to the path and renames over it, which replaced a symlinked lock with a detached regular file (uv itself writes through the link). Symlinked rewrite targets are now refused before any write — redirect_symlinked_file_unsupported (hosted, before the ledger is written), pypi_uv_symlink_unsupported, pypi_lock_symlink_unsupported — matching the hosted revert side, which already refused them.

Unwired-revert guard (Bugbot finding 1 and three adjacent holes)

• Skipped under --preserve-state (keep_artifact): nothing is deleted, so nothing needs protecting (npm's precedent).
• A project root that cannot be listed, or a lock whose symlink target cannot be read, refuses fail-closed instead of treating "no locks enumerated" as "no lock references it" and deleting the wheel (reproduced with an execute-only root).
• -r / --requirement includes of requirements.txt are probed, and repair / the orphan sweep see include-hosted pins too.
• A genuinely orphaned artifact directory (lock gone, or a flavor the dispatch does not know) is reclaimed instead of failing forever.

Robustness

• Hosted scan / get, repair and ledger-less rollback read candidate lockfiles through the FIFO-safe reader; a named pipe at a lockfile name no longer wedges the command in open(2).
• Harness: uv export is read from stdout (--output-file only exists from 0.4.7, so the 0.4.1 export lane had been recording a harness error as "unavailable"); uv 0.1.44 pinned to bracket the uv lock boundary; --render-doc-table regenerates the doc's results tables from results.json.

Probes that settled coverage questions (real binaries, evidence in the doc's Limits section)

• [[distribution]]-era vendored native stays refused: ≤ 0.2.6 cannot parse a relative path source at all; 0.2.17–0.2.34 install one under --frozen / plain sync but --locked rejects it, uv lock and plain sync absolutize it, resolution is CWD-relative, and hashes are unverified before 0.2.34.
• uv ≤ 0.1.23 has no local-path requirement grammar at all; the only spelling those binaries accept (name @ file:./…) breaks pip < 24.2, so the vendored requirements writer stays as is. --hash is silently ignored through 0.1.29; --require-hashes exists from 0.1.32; in-file hashes are verified by default only from 0.5.x. Documented.
• Every [[package]] release writes the bare [package.metadata] header even for dev-only projects, so the metadata no-op gate never fires for them.

Validation

• cargo clippy --workspace --all-features -- -D warnings clean; cargo test --workspace --no-fail-fast green (one pre-existing timing test in the update notifier flaked under the concurrent matrix load and passes on rerun).
• Real matrix (scripts/backtest-uv.py) on 42 uv releases spanning every family 0.0 through 0.12 at source commit 17d0dcb: 1403 installed-byte comparisons, 1086 lock-preservation checks. Every mismatch (18) and lock change (32) is the documented uv < 0.5.6 boundary in the transitive and constraints variant lanes — from 0.5.16 every shape passes in both modes with the lock byte-unchanged; on 0.2.35–0.5.3 --frozen installs the patch everywhere and the advisories name the limitation. Core lanes: 0 mismatches, 0 lock changes. The last commit after that run adds only the constraints advisory and the regenerated docs.
• New project-variant lane covers [tool.uv] dev-dependencies, [dependency-groups], dependencies + extra duplicate, constraint-dependencies, and a transitive dependency on every [[package]] binary, hosted and vendored, with --frozen, --locked and plain installs.

Hosted mode covers requirements from uv 0.0.5 and native uv.lock from 0.1.45 (the first release whose uv lock writes one); vendored native covers every [[package]] release (uv ≥ 0.2.35, with the ≥ 0.5.6 advisories for overrides and constraints); vendored requirements cover uv ≥ 0.1.24.

🤖 Generated with Claude Code

[cursor]

Cursor Bugbot has reviewed your changes using high effort and found 2 potential issues.

Bugbot Autofix prepared a fix for the issue found in the latest run.

• ✅ Fixed: Preserve-state fires empty-wiring refusal
  • Added !keep_artifact condition to empty-wiring guard in revert_pypi_opts to allow preserve-state mode to succeed without deletion refusal.

Or push these changes by commenting:

@cursor push c7e80e8475

Preview (c7e80e8475)

diff --git a/crates/socket-patch-core/src/vendor/pypi.rs b/crates/socket-patch-core/src/vendor/pypi.rs
--- a/crates/socket-patch-core/src/vendor/pypi.rs
+++ b/crates/socket-patch-core/src/vendor/pypi.rs
@@ -960,7 +960,7 @@
         dry_run,
         keep_artifact,
     } = opts;
-    if entry.wiring.is_empty() {
+    if !keep_artifact && entry.wiring.is_empty() {
         let uuid_dir_rel = vendor_uuid_dir_rel("pypi", &entry.uuid)
             .unwrap_or_else(|| format!(".socket/vendor/pypi/{:?}", entry.uuid));
         if let Some(blocked) =
@@ -3824,7 +3824,7 @@
         }
     }
 
-    // ───────────── uv guard failures surfaced through the orchestrator ────────────���
+    // ───────────── uv guard failures surfaced through the orchestrator ──────���──────
 
     #[tokio::test]
     async fn uv_lock_parse_failure_refuses_through_orchestrator() {

_{You can send follow-ups to the cloud agent here.}

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

^{Reviewed by Cursor Bugbot for commit 3726770. Configure here.}