[mozzius]

Summary

Fixes #2019.

layout.computedFlexBasis persists across YGNodeCalculateLayout calls and is only cleared by markDirtyAndPropagate on that exact node. The resolved-flex-basis branch of computeFlexBasisForChild keeps whatever value is already stored, which is what lets a max-content measurement made earlier in the same layout survive into the definite-size pass (and is why flex: 1 inside an auto-height column means "content height"). The guard only checks whether a value exists, not which layout produced it.

When a later layout at a different root size hits an ancestor's measurement cache during the max-content pass, that subtree is skipped, the child's computedFlexBasis is never refreshed, and the definite pass retains a size measured against the previous available size. In React Native this shows up as a flex: 1 Text that keeps its landscape height after rotating back to portrait (react/react-native#58294).

Fix

Record whether a stored computedFlexBasis came from measuring the child's content (computedFlexBasisIsMeasured). In the resolved branch, a stored value that is measured and stamped with a previous generation is treated as stale: the child takes the measure path again, which does exactly what the skipped max-content pass would have done and normally hits the child's own measurement cache. A retained resolved basis is unchanged, so "definite from the start" containers still resolve flex: 1 to a basis of 0 as before, and nothing changes within a single layout.

Clearing the value instead (or enabling WebFlexBasis) is not a fix: the resolved branch then writes 0 with no measurement to follow and the child collapses to 0.

Test plan

• New test in tests/YGRelayoutTest.cpp: an 8-node tree (scroll container, flex: 1 column with a fixed-height sibling, two plain wrappers, a flex: 1 node with a text-like measure function) laid out at 400x800, 800x400, 400x800. Asserts the measured node returns to its first height. Fails on main, passes with this change.
• ./unit_tests Debug: all 843 tests pass. On main the new test fails (portraitHeight is 100, the relaid-out node is 60) and the other 842 pass.
• Standalone repro (https://github.com/mozzius/rotation-text-repro, yoga-harness/run.sh) built against this branch: two and three wrapper levels now return to their first height (they stay at the landscape height on main); one wrapper level passes on both.

[metkm]

Hope this lands soon!. This is an annoying bug

[simongt]

We hit this in a production app on React Native 0.88.0 (iOS), and #2020 fixes it for us.

Our app is portrait only, except for a full-screen photo viewer (a transparent Modal with supportedOrientations). Rotating there and back leaves the screen behind it with a flex: 1 description that keeps its landscape height, and the rows below overlap it. That made it look like a Modal bug at first, but a plain rotation round trip does the same thing.

Independent repro, core React Native only (no react-native-screens), Release builds on an iPhone 17 Pro simulator, iOS 26.5: launch in landscape, rotate to portrait, rotate back to landscape.

Build	Back in landscape: flex: 1 item	Row below it, y
0.88.0 prebuilt	826 x 97.3 (its portrait height)	121.3
0.86.3 prebuilt	same	121
0.88.0 from source + #2020	826 x 49.3	73.3

A native probe on the stock build shows the path from #2019: the list's max-content measurement is a cache hit (826 x nan -> 826 x 122.7), then computeFlexBasisForChild keeps the item's computedFlexBasis of 97.3 from the previous generation. With #2020 applied, the same steps lay out correctly, and so does the Modal variant.

Before, stock 0.88.0:

bug-0.88.0-stock-round-trip-no-modal.mp4

After, with #2020 applied (the round trip, then the Modal variant):

fix-yoga-2020-round-trip.mp4