[dinwwwh]

Requested by dinwwwh · project thread

#88 made 3.1 to 3.0 about 6 to 7% slower on $ref-heavy docs. Wall time on the bench doc went from 15.1 to 16.1 ms, and on Discord from 19.1 to 20.3 ms. CodSpeed's simulation mode doesn't catch it. The cause: convertSchemaRef resolved and checked the target of every kept schema $ref and added it to a WeakSet. Only a $ref beside the null branch of a two-branch anyOf ever reads that result.

This PR moves the check into mergeNullable, so it runs only for those pairs:

• shared.ts exports find, the uncached pointer lookup that resolvePointer already wraps.
• NULL_REJECTING_REFS and the check in convertSchemaRef are removed.
• downgradeV30 passes root to mergeNullable. When the other branch is a kept $ref, mergeNullable checks the target in the input with rejectsNull(find(root, ref)). The input is always complete, so a cycle can't hand it a half-converted target.

Output doesn't change. All 256 tests and both snapshots (oRPC and corpus) pass with no edits. A separate check found byte-identical output to #88 on 41 real docs and 5,403 fuzz cases, with speed back to the pre-#88 level (15.4 ms and 19.0 ms). Locally, vitest bench puts this branch ahead of main on "generated api, 100 resources" in both runs. pnpm run check passes.

---

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

✅ 21 untouched benchmarks

---

_{Comparing claude/project-thread-o6uyg9 (133ed5f) with main (0321f9d)}

[pullfrog]

✅ No new issues found.

Reviewed changes — the perf follow-up to #88 that defers the null-rejecting $ref check from conversion time to mergeNullable.

• shared.ts — exports find, the uncached pointer lookup resolvePointer wraps, with a doc comment. No logic change.
• v3.1-to-v3.0.ts — removes NULL_REJECTING_REFS and its fill in convertSchemaRef; mergeNullable now takes root and, for a kept $ref branch, checks rejectsNull(find(root, ref)).
• Equivalence — old resolvePointer(ref, ctx) resolved against the same immutable input root that downgradeV30 now passes in, and $ref branches reaching the merge always originate from convertSchemaRef, so the WeakSet identity check and the getRef check cover the same branches. The full downgrader suite (252 tests) passes, including the existing kept-$ref and cyclic-schema cases.

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