react

Summary Fixes #37610. can return instead of or on alternating renders. The sibling traversal compares Fibers by identity, but DOM nodes and Fragment boundaries can refer to the alternate of the Fiber being visited. Recognize the alternate in both the target and boundary comparisons, as the nearby containment and common-ancestor checks already do. The existing traversal order and boundary checks remain in place, including rejection when DOM order disagrees with the Fiber tree. No public API signatures change. Merged upstream () in . Resolved the test insertion conflict by preserving both this PR's three sibling-position regressions and the upstream root-container/ancestor test from #37613. The PR diff remains limited to two files. How did you test this change? All results below are local, using Node v20.19.0, Yarn 1.22.22, and (passed). They were rerun on the merged tree, not copied from the original PR's results. Red/green regression check Temporarily removed only the four alternate comparisons to restore upstream behavior, then ran: Each baseline run had 2 expected failures and 1 passing control: repeated-render and keyed-reordering checks expected /, but received ; the DOM/Fiber-order disagreement control passed. After restoring the fix, both exact commands passed all 3 selected tests. Related suites and compatibility During the merge, relative to the old PR head reported whitespace in incoming upstream compiler snapshots. Its diagnostics were byte-for-byte identical to before the merge commit. Those unrelated upstream files were not edited; the actual PR diff check passes. GitHub CI For commit , GitHub reports the PR as mergeable (the conflict is resolved), but merge state is . Meta CLA Check: passed. (Runtime) Build and Test: action_required, 0 jobs started. (Shared) Lint: action_required, 0 jobs started. (Runtime) ESLint Plugin E2E: action_required, 0 jobs started. These are not failing test jobs or flaky-test results: the workflows have not executed. Maintainer action is required before upstream CI can be validated; the contributing account has no write/maintain/admin permission on the upstream repository. GitHub CI is not yet verified, and local passes above must not be read as CI passes. AI assistance: OpenAI Codex assisted with implementation, regression tests, review, conflict resolution, and validation.

Summary Fixes #37712 Store pending transition types per lane instead of sharing one list across the root Merge and deduplicate types only for the committing lanes and their entanglements before clearing finished lanes Clear types from completed or abandoned lanes so later transitions cannot inherit stale types Add coverage for unrelated Suspense reveals, independent and entangled commits, lane reuse, and async and nested transitions How did you test this change? Before the implementation change, three regression cases failed because unrelated commits consumed types, independent transitions mixed types, and abandoned transitions leaked types Local checks: One parallel validation run terminated a ReactDOMFloat worker with SIGSEGV under Node v24.19.0 on macOS, with the crash report pointing to V8 garbage collection The standalone rerun and full rerun passed without code changes, although the full rerun emitted a worker shutdown warning

Summary Babel can reuse a plugin instance across transforms. The Rust compiler's Program re-entry guard is currently shared by the instance, so a second source file can be skipped after the first one runs. Keep the guard scoped to each Babel file. Babel can still revisit a replaced Program without compiling it twice. How did you test this change? Built the native Rust addon and ran the new Node test through Babel. It uses one plugin instance for two files. Before the patch only the first received a compiler runtime import. Both receive it after the patch. The Rust CI workflow now runs the regression after building the native plugin. All 1,849 Rust snapshot tests passed. Prettier and passed. AI assistance was used during development.