feat(oxc): add vendored anti-slop capability and shrink-only baselines

[norvalbv]

Closes SC-1676.

Depends on #400. This PR is based on the Oxfmt self-host migration so the overlapping package,
generated-distribution, and toolchain-decision changes are reviewed as one coherent stack. Merge
#400 first; this PR can then be retargeted to main without changing its anti-slop commit.

Summary

• vendors the complete 15-rule dmmulroy/anti-slop production source at commit
  446268e5d15baa968eaec669ff65358d36ae6259 under its MIT license
• adds opt-in devkit init --anti-slop installation; it implies Devkit's exact-pinned Oxc
  capability and installs a self-contained @oxlint/plugins@1.78.0 runtime without changing the
  consumer dependency graph
• composes the namespaced rules through Devkit's managed Oxlint base, leaving the repository's
  ordinary Oxlint config in control of native/custom rules, off/warn/error, rule options, and
  scoped overrides
• adds explicit create, read-only check, read-only inspect, and shrink-only prune baseline
  operations
• fails closed unless a bounded runtime sentinel proves the repository root loads the exact managed
  Oxc base and registers the managed plugin; nested configs cannot shadow that proved chain
• wires init, interactive selection, upgrade offers, doctor repair, clean/removal, help, docs, and
  emitted dist/ assets

Why this boundary

Anti-slop is a direct fit for Oxc's JS-plugin surface, so this PR moves that policy onto the pinned
Oxc runtime. It does not turn this into a broader toolchain migration: Biome remains Devkit's
formatter/linter where selected, ESLint remains responsible for existing cross-file/filesystem
topology plugins, and tsc --noEmit remains the type checker. The updated
oxc-toolchain-migration decision note records that boundary so later work can change it only with
its own parity and performance evidence.

Baseline and configuration contract

• all rules default to error and are namespaced as anti-slop/<rule>
• .anti-slop-baseline.json has no timestamp and is sorted deterministically
• fingerprints are SHA-256 over rule ID, POSIX repository-relative path, normalized diagnostic,
  and normalized reported source line; absolute checkout paths and line numbers are excluded
• identical findings carry counts, so an added third instance above a baselined count of two is new
• normal checks never write; new errors fail while new warnings remain visible but non-blocking
• create refuses to replace an existing baseline without explicit --force
• prune can only remove absent debt or reduce counts, refuses new errors, and preserves all
  baseline entries outside a requested path scope
• paths must be literal existing files/directories inside the repository; external/missing paths are
  rejected
• baseline scans disable nested config discovery and use the repository root's overrides for
  scoped policy; managed/agent directories are excluded from adoption debt
• devkit clean removes only managed plugin state and deliberately retains the consumer's baseline

Packaging and provenance

The upstream production tree is byte-for-byte unchanged. A digest regression test pins its 19-file
tree, the build emits JavaScript under dist/anti-slop, and the installer copies the plugin API to
the capability-specific, trackable plugin/oxlint-plugins-api/ path. The installed copy uses a
package import map, so consumers need no direct dependency and common node_modules / vendor
ignores cannot omit runtime bytes. The managed manifest records upstream commit, plugin API version,
complete rule registry, and plugin/config/probe digests. Doctor and baseline operations verify
provenance, exact Oxc-base bytes, plugin completeness, and runtime composition. Replacement is
lock-serialized and rollback-safe; interrupted init is recovered from on-disk manifests on retry.

Evidence

• bun run format
• bun run typecheck
• bun run lint
• bun run lint:structure
• bun run build
• focused anti-slop/Oxc/init/clean suite: 110/110 passed across 9 files
• packed anti-slop capability E2E: 6/6 passed, including all 15 diagnostics; deterministic baselines across
  two checkout roots; missing baseline; create/check/new violation/fix; warning/off/scoped
  overrides; nested-config isolation; full-chain fail-closed behavior; interrupted-init recovery;
  path-scoped prune; doctor; clean with baseline retention; and package-mode git add/commit/clone
  with both node_modules and vendor/ ignored, followed by doctor/check without reinstalling
• packed Oxc lifecycle E2E: 1/1 passed against the composed managed base
• full repository suite: 3,834 passed, 5 skipped; 219 files passed and 1 skipped (220 total)
• vendored source compared byte-for-byte with the pinned upstream checkout
• generated package verified to include compiled plugin/lifecycle assets and no raw TypeScript plugin
  files
• correctness review passed all four lenses after adversarial repros covering config comments,
  direct-fragment composition, nested shadowing, managed-base drift, concurrent replacement,
  interrupted install/clean, scoped check/prune, symlinks, and ignored delivery layouts; the final
  ship review additionally caught and fixed scoped create --force baseline replacement, with an
  E2E proving out-of-scope debt is preserved. A second review cycle caught stale filesystem-derived
  anti-slop activation and a non-exercising runtime sentinel; the Oxc manifest now records the
  selected integration explicitly, every init/doctor/install/remove caller supplies it, and the
  sentinel must observe an actual managed-rule diagnostic while ordinary lint suppresses the
  intentional probe through a nested managed config. A third cycle moved the paired anti-slop
  removal/Oxc rewrite under one lock; it now proves the base can be unwired before deleting plugin
  bytes, with lock-held and runtime-failure preservation regressions.
• final GitNexus change detection found the expected selection/init/upgrade/clean execution flows;
  the shared defaultSelection hub makes the graph risk high, and its direct dependents plus the full
  suite were validated

Runtime sanity sample

This is not a replacement for the earlier Frink-scale migration benchmark. On an Apple M4 / 24 GiB
Mac, Node 22.20.0, and Oxlint 1.78.0, the final sentinel-inclusive wrapper over a copied Devkit
authored-code surface (cli/ plus gate-engine/) produced a deterministic baseline of 1,613
findings / 1,486 fingerprints in 0.65 s wall with 223 MB maximum RSS. Seven subsequent
baseline-aware checks had a 0.60 s median wall time and 187 MB median maximum RSS
(/usr/bin/time -l); the baseline SHA-256 remained
ac7fb9b2119c035d795f8f3448289170c1c5069b9af00b0b3870b8f22c45941d. This bounds the final
wrapper/adoption overhead; the saved Frink benchmark/decision evidence remains the source for
cross-tool migration claims.

Known limits

• Oxc's JS-plugin API remains alpha, so the exact Oxlint/plugin API pins and packed all-rule fixture
  are the compatibility boundary for upgrades.
• These rules are syntax/scope based and do not replace TypeScript checking.
• A consumer root Oxlint config must retain ./.devkit/oxc/oxlint.base.json in its extends chain;
  doctor reports/repairs managed-side drift, while a consumer-owned missing root pointer requires an
  explicit config edit.
• Baseline operations intentionally disable nested config discovery; use root-level Oxlint
  overrides for path-specific anti-slop policy.
• Identical findings within one file are fungible counted debt units. Fixing one while adding an
  indistinguishable occurrence keeps that fingerprint's count flat; this is intentional so
  unrelated line shifts do not explode an adopted repository's baseline. The ship gate's
  count-offset finding is waived with this explicit non-growth rationale.

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on base/target branches other than the default branch.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro Plus

Run ID: da6914d1-a1bd-4210-bea1-d7c66c2449a8

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}