Tracking: revert every vendored backend through one record-revert engine instead of nine hand-written mechanisms #989

[agent] Filed by the scheduled architecture audit routine (ecosystems and formats). Register: register comment.

Kind: tracking. Source: review §2.1, Part 5.3 and 5.8; register E24.

Problem

Every vendored backend's revert_*_opts re-implements the same envelope by hand, and the record restore under it is done about nine ways (Part 5.3). On 9c43dfc:

1. The finish step is copied 12 times. It runs: dry-run return → drift-keep → --preserve-state return → remove_tree_and_prune(uuid_dir, .socket). The copies:

• bun_binary.rs#L528-L561
• bun_lock.rs#L920-L950
• composer_lock.rs#L586-L612
• gem.rs#L1242-L1271
• maven_repo.rs#L658-L691
• npm_lock.rs#L800-L876
• nuget_feed.rs#L766-L808
• pnpm_lock.rs#L1004-L1034
• pypi.rs#L1625-L1690
• vlt_lock.rs#L1908-L2015
• yarn_berry_lock.rs#L885-L917
• yarn_classic_lock.rs#L540-L571

The copies use four keep policies, which CLI_CONTRACT.md documents per backend:

• Unconditional drift-keep: gem, npm, pnpm, bun, bun binary.
• Keep only while a live file names the uuid dir (any_live_file_references): composer, maven legacy, nuget.
• A post-revert "lock still mentions the uuid" refusal, vendor_lock_still_wired_revert_blocked: yarn classic and berry.
• Drift or vendor_revert_residual_reference keeps the artifact, and a removal failure is only a warning (vendor_artifact_remove_failed): pypi. Every other backend makes a removal failure an error.

2. The record loop is copied 4 times. The uuid gate, a reverse for w in entry.wiring.iter().rev(), match w.kind, an "unrecognized wiring kind {:?}; fragment left alone" warning, and Ok(true) / Ok(false) → vendor_lock_entry_drifted / Err → fail. The copies:

• composer_lock.rs#L558-L585
• gem.rs#L1194-L1240
• maven_repo.rs#L623-L657
• nuget_feed.rs#L725-L765

3. Only poetry and pdm use the one shared splice helper:

• common.rs#L696-L830

Only that helper has an all-or-nothing mode for coupled records (revert_lock_fragment_splice_atomic).

4. Other differences:

• requirements.txt speaks its own drift code, vendor_revert_line_drifted, which CLI_CONTRACT.md doesn't document.
• cargo writes a cargo_lock_entry wiring record that its revert never reads (cargo.rs#L1219-L1234), because it restores from the typed entry.lock (cargo.rs#L1978).

Symptoms

• Vendored gem revert restores Gemfile.lock but leaves the Gemfile path: line when that line carries a trailing comment, and no re-run can finish it #988: the gem forward recognizer is a substring and the revert recognizer an exact line, and the coupled records are reverted non-atomically, so the revert is half-applied and drift-keeps forever.
• Vendored requirements.txt can't be reverted once the user touches the # socket-patch vendor: comment: vendor --revert drift-keeps forever (exit 0), and the suggested "re-vendor" remedy is a no-op or adds a duplicate line #977: requirements.txt has the same forward/revert recognizer split.
• Vendored yarn classic rollback can't undo a block yarn has merged with another range (left-pad@^1.1.0, left-pad@^1.3.0:): it reports the block as gone, exits 1 forever and leaves the lock wired #692: yarn classic can't revert a merged block.
• Vendored NuGet on a core.autocrlf checkout: vendor --revert / remove / rollback revert packages.lock.json but leave nuget.config wired, so every restore fails NU1403 (vendor --revert exits 0) #537: NuGet revert on a core.autocrlf checkout.
• Fix uv vendored revert half-reverting after a relock (#806, #821) #822 (fixed) was the same half-revert class in uv.

Impact: every backend re-decides the drift-keep, atomicity and "is this still our wiring" rules, so a fix lands in one backend at a time. The revert code in the non-npm backends alone is about 3.5K lines.

Target design

• One revert_entry(entry, opts, backend) engine in vendor/revert.rs that owns the envelope:
  • the uuid gate;
  • the pre-revert refusals each backend supplies;
  • a plan pass over records in reverse order, where each record yields Restore(file, new_text), Converged, Drifted(detail) or Unknown;
  • an all-or-nothing write for records that a backend marks as coupled;
  • one finish step with one documented KeepPolicy.
• Each backend supplies only plan_record(&WiringRecord, &ReadFn), plus the forward-path recognizer it already has, so forward and revert share one "is ours" predicate.
• Legacy ledger kinds are adapted to current records at load, never by a second revert path.

Checklist (in order)

• 1. Finish every vendored revert through one shared helper with an explicit keep policy, deleting the 12 copies (child 1: Finish every vendored revert through one shared helper with an explicit keep policy instead of 12 copied finish blocks #990; no behavior change).
• 2. Share the record loop and the unknown-kind warning (composer, gem, maven legacy, nuget) through a RecordRevert { Done, Drifted, FileMissing } driver with an atomic option. Blocked by 1.
• 3. Move poetry, pdm and pipenv onto the driver (pipenv's canonical rewrite becomes a JSON-value record). Overlaps Share the single-lock wire and revert envelope across the Poetry, PDM and Pipenv vendored backends #937.
• 4. Move requirements.txt and the pypi_lock three-way merge onto the driver, with the forward-path recognizer (fixes Vendored requirements.txt can't be reverted once the user touches the # socket-patch vendor: comment: vendor --revert drift-keeps forever (exit 0), and the suggested "re-vendor" remedy is a no-op or adds a duplicate line #977). The vendor_revert_line_drifted decision is whether to keep it as an alias or fold it into vendor_lock_entry_drifted.
• 5. npm family: express the seven lock reverts as record plans after Tracking: drive the seven npm-family vendor backends through one generic driver #920's generic driver.
• 6. Decide whether to unify the four keep policies into one, a contract change filed as a decision when this step is reached.
• 7. Delete ledger_snapshots once no backend records whole-file snapshots (with E27).

Dependencies

• Related: Tracking: dispatch vendored backends through one per-ecosystem table instead of string matches in core and the CLI #959 / Route vendored revert and in-use dispatch through one core VendorBackend enum instead of CLI string matches #960 (backend dispatch), Tracking: drive the seven npm-family vendor backends through one generic driver #920 (npm driver), Share the single-lock wire and revert envelope across the Poetry, PDM and Pipenv vendored backends #937 (PyPI envelope).
• Vendored gem revert restores Gemfile.lock but leaves the Gemfile path: line when that line carries a trailing comment, and no re-run can finish it #988 can land independently first.

Consolidated work — backlog review, 2026-10-08

The following standalone issues are now tracked here. Their closure consolidates scheduling; it does not mean their implementation is complete. Original reports and discussion remain linked below.

#990: Finish every vendored revert through one shared helper with an explicit keep policy instead of 12 copied finish blocks

• Track the remaining scope from Finish every vendored revert through one shared helper with an explicit keep policy instead of 12 copied finish blocks #990 in this issue. Explicit shared revert-finish child; retain its per-backend keep-policy table under the parent rather than another issue.