[hryhoriiK97]

Summary

#59290 added a virtual asLayoutable() to ShadowNode, right after completeClone. Every virtual method declared after it (appendChild, replaceChild, layoutTree, measureContent, layout, ...) moved one vtable slot down.

Native code built against the old headers, for example a Fabric component from a prebuilt library, still has the old vtable layout, so a call from the core can land on a different method. In my repro, the core asks such a component for asLayoutable() and the call lands in its appendChild:

SIGSEGV
ShadowNode::appendChild
YogaLayoutableShadowNode::appendChild
ShadowView::ShadowView
FabricMountingManager::maybePreallocateShadowNode
UIManager::createNode

This may also explain a report we got from an app on a nightly: the JS bundle loads, nothing crashes and nothing is logged, but the first screen never appears. We have not confirmed that yet.

This PR keeps the speed-up from #59290 without the vtable change:

• asLayoutable() is removed, so the ShadowNode vtable is the same as before Use a virtual accessor instead of a dynamic_cast when building ShadowView layout metrics #59290.
• A new trait, LayoutableKind, is set by both LayoutableShadowNode constructors. They live in the core library, so subclasses built against older headers get the trait too.
• ShadowView checks the trait: with it, a static_cast; without it, empty layout metrics, as asLayoutable() returned nullptr for such nodes.

There is no behaviour or performance change compared with main: every node gets the same layout metrics as with asLayoutable(), still without a dynamic_cast. Only the vtable change is undone.

Changelog:

[GENERAL] [FIXED] - Keep the ShadowNode vtable layout compatible with native code built against older headers

Test Plan:

• Repro on an Android emulator: RNTester with its own codegen components (react_codegen_AppSpecs, incl. RNTReportFullyDrawnView) compiled against the pre-Use a virtual accessor instead of a dynamic_cast when building ShadowView layout metrics #59290 headers, the rest against main.
  • main: SIGSEGV 2 s after launch, with the backtrace above.
  • this PR: launches, opens FlatList → Basic, scrolls to item ~580, no crash.
• Vtable of a custom ConcreteViewShadowNode subclass, dumped with -fdump-vtable-layouts: identical to the one before Use a virtual accessor instead of a dynamic_cast when building ShadowView layout metrics #59290.
• yarn fantom, full suite, with the tester rebuilt from this change: 3316 tests pass. The 60 failures are all snapshot printing in my checkout's Jest setup (React elements printed as plain objects instead of JSX), including Fantom's own expect-itest.js; none is about layout.
• cxx-api snapshots updated; clang-format clean.

🤖 Generated with Claude Code

[javache]

This should go in BaseTraits

[javache]

React Native is not ABI stable - it's incorrect to assume you can existing binaries with newer versions of the framework.

[hryhoriiK97]

React Native is not ABI stable - it's incorrect to assume you can existing binaries with newer versions of the framework.

@javache , thanks for the review! For context, I found this because code built against the headers from before #59290 crashes on main: the asLayoutable() call lands in appendChild and it ends in a SIGSEGV. That's also why I set the trait in the constructor and not in BaseTraits: BaseTraits is inline in the header, so code built against older headers would never set it. What do you think we should do here?

[javache]

Code being built against previous headers is an unsupported use-case, and I wouldn't land this change just for the sake of that.

If you think modelling this as BaseTraits instead of asLayoutable, that's something we can merge, although I think it should go in the header too.

[hryhoriiK97]

@javache makes sense, thanks. Without that case there's no real gain over asLayoutable(), so I'll close this one 👍