CI: rollback tests assert codes, not sentences (#1089)

[Mikola Lysenko (mikolalysenko)]

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

Refs #1089 (child 4: rollback). #704 and #1027 have closed, so rollback's --json error is a {code, message} object and this child is unblocked.

What changes

• tests/covgap_commands_rollback.rs (3,568 lines, 49 tests, its own binary) is split into the rollback suite binary:
  • rollback/rollback_failure_paths.rs (29 tests) checks exit code, status, error.code, event errorCodes, warnings[] codes and disk state through common::envelope and rollback_json. Messages are read only for what they must name (an ecosystem, a path, a remedy). It has 0 .contains("…") asserts of 4+ words; the old file had 33.
  • rollback/rollback_render.rs (27 tests) holds the human output checks, one per message family. This includes the human halves of the old *_in_both_modes tests, which are now split into a JSON test and a render test. 25 sentence asserts remain here, all in the render suite, as the tracker's target design calls for.
  • rollback/rollback_fixture.rs holds the shared fixtures once: the vendored and hosted fixtures, the mock npm registry, the DirModeGuard, and a single macOS ChflagsGuard where there used to be two inline copies.
• Sentence matches replaced by codes: vendor_ledger_missing (ledger-less wired lock), vendor_revert_failed (unknown backend), vendor_state_write_failed (ledger save), hosted_restore_refused (per-pin refusal), hosted_write_failed (lockfile write, message match dropped) and manifest_invalid (message match dropped).
• Two human-only behaviour tests now run --json: the blob-hash directory squatter (already_original count) and the Bun orphaned store entry (not_installed empty, rolledBack names the purl).
• coverage_fix_rollback_ecosystem_scoped_hosted.rs moves unchanged into the same binary as rollback/rollback_ecosystem_scoped_hosted.rs. It had no sentence asserts.
• cli/envelope_helper_copies.rs: covgap_commands_rollback.rs leaves the pending-private-readers list, because its private parse_envelope/warning_codes copies are gone.
• docs/testing/bun-compatibility.md points at the new files.
• scripts/ci-test-durations.json (Windows shard timings only): the two removed binaries' entries are dropped, and rollback goes from 4.66s to 6.61s (their runtime summed; per-target compile time is counted separately).

CI cost

Workflows are unchanged. A lean PR run has the same ~28 Linux jobs before and after, at roughly the same job-minutes (~8 min wall). The Linux test shards link two fewer integration-test binaries, which saves a few seconds per shard. That is too small for a ci-perf label.

The moved tests include cfg(unix) and cfg(target_os = "macos") bodies. Locally, I compiled the suite with the macOS cfg flipped to unix, and again with every unix/macOS item compiled out (approximating Windows). Both built with no warnings. The one full-scope run on ab2e710 passed all 99 jobs, including the macOS and Windows test shards: https://github.com/SocketDev/socket-patch/actions/runs/38651659602. The follow-up commit 3c229e2 changes no platform-gated code.

Test evidence

• cargo test -p socket-patch-cli --all-features --test rollback: 119 passed. The vanished-manifest choreography test passed 5 runs in a row.
• cargo test -p socket-patch-cli --all-features --test cli envelope_helper: 11 passed (the private-reader ratchet).
• cargo fmt --all -- --check: clean.
• cargo clippy --locked --workspace --all-features -- -D warnings (CI's invocation): clean. With --test rollback, clippy reports nothing in the new files; the errors it shows are in shared modules and appear on main too (e.g. --test repair).
• CI on 3c229e2: ci-ok and clippy green, and every lean job passed. Cursor Bugbot's one finding (env leakage from the in-process tests) was a false positive: replied with the evidence and resolved; the re-review found no issues.
• python3 -m unittest discover -s scripts/tests: 347 OK, 6 skipped. This includes the shard planner test that reads ci-test-durations.json.

/code-review high

Fixed: the vanished-manifest test no longer tells its two interleavings apart by sentence (it now checks whether the message names manifest.json); a redundant negative sentence match is dropped; an inverted failure message is fixed; a shadowing IO_PURL local is renamed; stage_blob now uses common::write_blob; the human legacy-ledger test keeps ../evil.lock inside its own tempdir; the docs list both files; a malformed comment is fixed.

Not taken: move NpmRegistry / DirModeGuard / readonly_dir_enforced into tests/common. Near-identical copies live in the remove and get suites, so merging them is a cross-suite change outside this slice. This PR goes from two copies to one within the rollback binary.

Still open in #1089

Child 5 (covgap_commands_scan_mod.rs, covgap_commands_scan_hosted.rs, covgap_commands_get.rs) and child 6 (render suites and the remaining non-covgap files).

🤖 Generated with Claude Code

https://claude.ai/code/session_01HsgT23cgsoRcwB1reaDykA

---

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]

✅ 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 3c229e2. Configure here.}

[Mikola Lysenko (mikolalysenko)]

Labeled Ready for review at head 3c229e2.

• CI: 86/86 checks green (36 success, 50 skipped), no main-wide failures.
• Bugbot: reviewed 3c229e2, no new issues; its one thread is resolved.
• Test-only change (rollback covgap split into the rollback suite). No CHANGELOG.md change.
• Reviewer focus: that the JSON-path tests now assert error.code/warning codes rather than English sentences, with sentence checks confined to rollback_render.rs.

---

Generated by Claude Code