Make pnpm CoW e2e independent of the live patch catalog

[Mikola Lysenko (mikolalysenko)]

Problem

e2e (*, e2e_safety_pnpm) is a required leg on ubuntu (PR + merge queue), Windows and macOS (main/merge queue). It fetched the live minimist@1.2.2 patch with socket-patch get <uuid> and pinned that patch's after-hash. When production republished the patch on 2026-10-09 (#1293), 3 of its 4 tests failed in every run until the repin (#1301) landed:

• 20 failed e2e (ubuntu-latest, e2e_safety_pnpm) jobs between 12:00Z and 18:20Z (CI runs on PRs and merge_group). Examples: merge_group head entries pr-1247 run 37957804376 (job 113914790149) and pr-1288 run 37961831426 (job 113928486385).
• Failure signature: apply_in_a_does_not_mutate_b_or_store, pnpm_install_in_b_does_not_revert_a and apply_in_pnpm_project_emits_layout_note fail on a hash mismatch (043f04d1… vs 311f1e89…) or a missing layout note.
• Effect on the queue: from 15:49Z to 18:02Z no merge_group CI run passed. The bot evicted queue entries 9 times (Fix bun.lockb shared bundled pin being unmanageable (#1243) #1247 ×3, Skip the vlt compat matrix on ci.yml-only changes #1284 ×2, Retry CI toolchain installs past rustup download blips #1286 ×2, Wire vendored nuget.config through formats::nuget (#594) #1288 ×2). Unlike hosted-e2e, this leg has no HOSTED_E2E_DISABLED-style kill switch.

Root cause

The suite tests the copy-on-write defense against a real pnpm store. It does not test the patch catalog, but it still depended on the catalog's exact contents.

Fix

Keep the real pnpm install (hardlinked shared store). Instead of get, stage a synthetic manifest and after-hash blob under proj_a/.socket/: the patched bytes are the pristine index.js plus a marker line. Then run socket-patch apply --offline. This is the same pattern e2e_safety_cow.rs already uses. BEFORE_HASH stays pinned, because published npm tarballs are immutable. Every CoW, inode-identity, frozen-reinstall and layout-note assertion is unchanged. The layout-note test now also asserts that the file was actually patched.

The live catalog keeps its coverage in e2e_npm, e2e_hosted_production and e2e_vendored_production, which #1301 repinned.

Proof

• cargo test -p socket-patch-cli --test e2e_safety_pnpm -- --ignored passes 4/4 locally (pnpm 10, Linux).
• Ran the built test binary 20 more times in a loop: 0 failures.
• cargo clippy -p socket-patch-cli --test e2e_safety_pnpm -- -D warnings is clean, and rustfmt was run on the touched file only.

Where the tests run

No test was removed or moved. The same matrix rows still run them: e2e ubuntu on PRs, e2e-windows and e2e-macos. No workflow changes.

🤖 Generated with Claude Code

https://claude.ai/code/session_01UcDQZrWwmbg27Ut4sxrtLw

---

Generated by Claude Code

[Mikola Lysenko (mikolalysenko)]

bugbot run

---

Generated by Claude Code

[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 24c8b2c. Configure here.}

[Mikola Lysenko (mikolalysenko)]

ci-ok failed on 24c8b2c because the CI run 37976351039 was cancelled from outside the run at 19:47:48Z, not because a test failed. 53 jobs passed and 41 were cancelled mid-flight. No job failed, and no newer run exists on this SHA, so this is not a change-caused failure (most likely a runner-backlog cleanup). I re-ran the failed and cancelled jobs once.

---

Generated by Claude Code

[Mikola Lysenko (mikolalysenko)]

On head 8c3671e (main merged in), e2e (ubuntu-latest, e2e_gradle_discovery_build e2e_gradle_agent_build e2e_redirect_gradle_build, …) failed because its Depot runner was lost mid-test, not because of a test assertion. The log shows The runner has received a shutdown signal 5 seconds into the second suite, then Step canceled by GitHub. This PR changes only e2e_safety_pnpm.rs and touches no Gradle code. I'll re-run the failed jobs once when the rest of run 37998395296 finishes.

---

Generated by Claude Code