Yarn berry hosted pin can't cover a descriptor added after pinning: re-scan refuses with an impossible "dedupe" remedy, and rollback reports success but leaves a lock every `yarn install --immutable` rejects (YN0028)

[mikolalysenko]

[agent] Found by the scheduled Yarn Berry (2+) bug-hunt routine (ledger #305).

Summary

A hosted berry pin is a root resolutions selector for each descriptor that existed when you ran the scan (for example "lpx@npm:^1.0.0": "<socket url>"). If someone later adds the same package under a different range, the selector doesn't cover it. That happens when a workspace member adds "lpx": "1.0.0", or when a new dependency depends on lpx@~1.0.0. Yarn then locks a second, separate registry entry "lpx@npm:1.0.0" beside the Socket URL entry, and installs that copy unpatched. Three things go wrong after that:

1. A re-scan can't fix it. scan --mode hosted exits 0 and pins nothing (redirect_yarn_berry_ambiguous_entry): "yarn.lock holds 2 separate entries resolving lpx@1.0.0; run yarn install once to dedupe the lock, then re-run". Neither yarn install nor yarn dedupe merges the two entries, because the resolutions pin is what keeps them apart. The suggested fix can't work, and nothing else is offered.
2. Rollback breaks the lock. rollback (and remove) exit 0 with "Restored pkg:npm/lpx@1.0.0", but they write back two blocks with the same resolution ("lpx@npm:1.0.0" and "lpx@npm:^1.0.0", both resolution: "lpx@npm:1.0.0"). Yarn merges those into "lpx@npm:1.0.0, lpx@npm:^1.0.0", so every fresh yarn install --immutable fails with YN0028.
3. Lock-only vex on main attests not_affected while packages/b/node_modules/lpx is unpatched. Open PR Stop VEX attesting over yarn PnP, pnpm bundled and deno.lock copies #1033 already fixes this part (its berry same-lock rule; I checked its head 314038b: lock-only vex now drops the ref with the "stays UNPATCHED" warning). But Stop VEX attesting over yarn PnP, pnpm bundled and deno.lock copies #1033 says a re-run of scan "rewires every copy", and for berry hosted that isn't true (point 1). With Stop VEX attesting over yarn PnP, pnpm bundled and deno.lock copies #1033 merged, the project loses VEX and still has no way to re-pin.

Impact

This is a normal workflow: someone adds a dependency after the hosted pin. Afterwards one copy of the patched package installs unpatched, the scan reports nothing it can do (exit 0), and the documented undo (rollback) leaves a lockfile that CI's --immutable install rejects.

Repro (local registry + mock patch API; synthetic lpx@1.0.0)

mkdir -p proj/packages/b && cd proj
echo '{"name":"root","version":"1.0.0","private":true,"workspaces":["packages/*"],"dependencies":{"lpx":"^1.0.0"}}' > package.json
printf 'nodeLinker: node-modules\n' > .yarnrc.yml
yarn install
socket-patch scan --mode hosted --yes                # pins lpx@npm:^1.0.0 → socket url (resolutions + lock)
echo '{"name":"b","version":"1.0.0","dependencies":{"lpx":"1.0.0"}}' > packages/b/package.json
yarn install                                          # adds a separate "lpx@npm:1.0.0" registry entry
head -1 packages/b/node_modules/lpx/index.js          # unpatched
socket-patch scan --mode hosted --yes                # exit 0: "holds 2 separate entries … run `yarn install` once to dedupe"
yarn install && yarn dedupe && socket-patch scan --mode hosted --yes   # same refusal, still 2 entries
socket-patch rollback                                 # exit 0 "Restored …"
# fresh checkout of package.json/yarn.lock/.yarnrc.yml:
yarn install --immutable                              # YN0028 (lock would merge the two entries)

Expected vs actual

• Expected: a re-scan pins the new descriptor too. The rewriter already pins merged multi-descriptor entries by adding one resolutions selector per range, so it can add lpx@npm:1.0.0 and fold the registry entry into the URL entry. If it can't, it should refuse with a remedy that works (exit non-zero, or at least name the descriptor to remove). CLI_CONTRACT.md / docs/ecosystems.md describe a re-run of scan as the way to re-wire a project whose lock drifted.
• Expected: rollback leaves a lock that yarn install --immutable accepts (the bar every other berry rollback cell meets). If it can't produce one, it should fail loudly rather than report success.
• Actual: as in the summary.

OS × version

OS	yarn	re-scan dead end	rollback → YN0028	lock-only vex attests (main)
Linux	4.0.2	reproduces (2/2)	reproduces (2/2)	reproduces (2/2); dropped on PR #1033
Linux	4.18.1	reproduces (2/2)	reproduces (2/2)	reproduces (2/2); dropped on PR #1033
macOS / Windows	—	untested (the probe branches are blocked)

Tested on main 05ecc6e, and on PR #1033 head 314038b for the re-scan and VEX columns. Not bisected. Releases ≤4.0.0 pinned through an ::__archiveUrl= locator instead of resolutions.

Suspect code

• crates/socket-patch-core/src/patch/redirect/mod.rs:4141-4152: targets.len() > 1 gives redirect_yarn_berry_ambiguous_entry with the dedupe remedy, even when one of the two targets is the existing Socket URL entry and the other is a registry entry whose descriptor just needs another selector.
• crates/socket-patch-core/src/patch/redirect/upstream/npm.rs:546 (restore_berry): it restores the URL entry's keys as their own block next to an existing block with the same resolution:, and doesn't merge them the way yarn would.
• crates/socket-patch-core/src/vex/discover/yarn.rs:405-409: a plain registry entry is only resolved_elsewhere, not an unpatched_copy (PR Stop VEX attesting over yarn PnP, pnpm bundled and deno.lock copies #1033 changes this).

---

Backlog review — 2026-10-08

Priority: P1 → P2. Adding a Yarn descriptor after pinning creates a refusal/broken immutable install. Retain the recovery fix; the dangerous VEX portion was addressed separately.

[mikolalysenko]

[agent] Triage: priority:p1 (Yarn berry, npm-family). Not a duplicate. Point 3 (lock-only VEX) is addressed by open PR #1033, as the report notes. Points 1 (the ambiguous-entry refusal at redirect/mod.rs:4141) and 2 (restore_berry not merging blocks that share a resolution) have no open PR. #962 also ends in YN0028 after rollback, but its cause is a user yarn patch, so I'm not clustering them.

---

Generated by Claude Code

[mikolalysenko]

[agent] Re-triage on main ea09714 (after #1033 and #1035). Linux, yarn 4.0.2 and 4.18.1, same repro as the report (workspace member adds lpx: 1.0.0 after a hosted pin of lpx@npm:^1.0.0).

Point	4.0.2	4.18.1	Status
3. Installed-tree vex attests the unpatched member copy	not attested; warns patched_ref_unattributable ("that copy stays UNPATCHED")	same	fixed by #1033
2. rollback exit 0, then fresh --immutable YN0028	now exit 1, status: error, nothing written; fresh --immutable passes, because the lock still holds the pin	same	now fail-closed
1. Re-scan dead end	scan --mode hosted exit 0, redirect_yarn_berry_ambiguous_entry ("run yarn install once to dedupe the lock")	same	still reproduces

Point 1 is now a loop. The rollback refusal says: "reconcile the lockfiles (re-run socket-patch scan --mode hosted) or restore them from version control (git checkout -- package.json yarn.lock)". The re-scan refuses with the dedupe remedy, and yarn install/yarn dedupe don't merge the two entries. So the only way out is git checkout, which also throws away the member's new dependency. The member copy stays unpatched, and nothing in socket-patch can re-pin it or unwind it.

---

Generated by Claude Code