mirror of
https://github.com/stablyai/orca.git
synced 2026-09-28 16:02:45 +00:00
m4air
908db722f7
React #185 crashes in v1.4.199 (reports 3efb42f1, 74dc8068) ship a crash bundle with no `react_commit_cascade` breadcrumb at all, so triage falls back to the boundary that happened to catch the throw — a bystander the payload itself flags as unreliable. The cascade observer only counted a commit when `root.pendingLanes` still held SyncLane/InputContinuousLane/DefaultLane. Verified against react-dom 19.2.8: `onCommitFiberRoot` is invoked BEFORE `0 !== (pendingEffectsLanes & 3) && flushPendingEffects()`, and React reads `root.pendingLanes` for `nestedUpdateCount` only after. A store-subscription loop schedules its re-render from a passive effect, so every commit samples 0 while React counts it and throws — measured: all 55 commits report 0. Count re-entrancy beside the lanes: a commit that arrives before the stack unwinds to a microtask checkpoint is inside React's own synchronous `flushSyncWorkAcrossRoots_impl` loop. The crumb ships `evidence` ('lanes' | 'reentrant' | 'mixed') and `laneCommits` so a measurement is never readable as an inference, and `pendingLanes` stays raw. A run with no lane evidence is dropped at its span's microtask checkpoint, which both bounds the inferred count to one synchronous span and stops `cascadeRoot` pinning a root unmounted mid-span through an idle window. This does not fix the loop. It makes the next occurrence name its driver.