[dinwwwh]

🤖 Generated with Claude Code

https://claude.ai/code/session_01VVaw7RYWCsn8myBE4Rncuq

A follow-up to #365, from a Stryker run on it.

• leadsBack drops its bare-key test and follows each $ref by itself. 3.0 output never has a $ref with siblings: there are 0 among the outputs of 46k fuzz docs. The loop is also shorter now.
• Tests cover two cases:
  • an object cycle in a pair's target. fix(downgrader): keep a null pair only when its $ref would loop back to it in 3.0 #365 dropped this case with the input walk, and it was the only test that killed an enumUnionType mutant.
  • a pair whose $ref target moves in 3.0 (into a type list's object branch). There the chain leads nowhere, and the old || mutant would have thrown.

Checks

• Tests pass with 100% branch coverage.
• Output is identical to main on the c1, c2, kn1 and kt1 fuzz dumps (46k docs).
• The loop fuzzer still finds 0 bare loops.
• Stryker: 78 survivors, down from 81 on main. The 3 new survivors left from fix(downgrader): keep a null pair only when its $ref would loop back to it in 3.0 #365 are equivalent: the initial value of collapsedPairs, and its length > 0 guard, which only skips work.

---

Generated by Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[codspeed]

Merging this PR will not alter performance

✅ 29 untouched benchmarks
⏩ 2 skipped benchmarks¹

---

_{Comparing claude/project-thread-rixcs1 (587947c) with main (437ae2d)}

Footnotes

1. 2 benchmarks were skipped, so the baseline results were used instead. If they were deleted from the codebase, click here and archive them to remove them from the performance reports. ↩

[pullfrog]

✅ No new issues found.

Reviewed changes

• leadsBack follows any $ref (packages/downgrader/src/v3.1-to-v3.0.ts:1741) — the walk drops its bare-record guard and node-identity seen set for a Set<string> of visited refs, reading getRef(schema) directly. Verified behavior-preserving: only next === schema returns true, the new walk is a strict superset of the old, and no output node can carry a $ref with siblings, so the extra reachability is unreachable (see below).
• Two edge tests (packages/downgrader/tests/v3.1-to-v3.0.test.ts:1406) — an object cycle in a pair's target converts while the pair stays a bare $ref (restores the case #365 dropped, which killed an enumUnionType mutant), and a pair whose target moves in 3.0 (into a type list's object branch) leads nowhere.

The bare-$ref invariant the new code relies on holds structurally, not just empirically. Every subschema-bearing schema keyword is either converted (allOf/anyOf/oneOf/not/properties/items/additionalProperties) or removed/renamed by REMOVED_KEYWORDS and the V30_SCHEMA_KEYWORDS sweep, and every converted $ref is emitted bare: convertSchemaRef and refOr build { $ref }, finishSchema moves a $ref plus siblings into allOf, and convertTaggedBranch routes through convertSchemaRef. A collapsed pair is always a one-key { $ref } when leadsBack runs, so the dropped guard cannot change its result; old and new also agree on dangles (getRef(undefined) is undefined). Tests pass locally (137 passed).

  ^{｜ View workflow run ｜ Using deepseek-v4.1-flash (free via Pullfrog for OSS) ｜ 𝕏}