Adding combined isPending breaks atomic publication

[GabbeV]

Describe the bug

Two async memos derive from the same signal. Reading isPending(() => [fast(), copy()]), where copy is a synchronous memo wrapping the slow result, makes Fast display 1 while Slow still displays 0. Pending reports false during this interval.

Removing the pending read makes both values update together.

Your Example Website or App

https://s.olid.uk/id/P4Zyb5IlR_Gmz8H0dvcBUQ

Steps to Reproduce the Bug or Issue

1. Wait for initialization.
2. Click Run.
3. Fast becomes 1 while Slow remains 0 and Pending reports false.
4. After one second, Slow becomes 1.

Expected behavior

The pending read should not cause a partially published update. Both values should update together, with Pending true until they are ready.

Screenshots or Videos

No response

Platform

Verified with compiled JSX in jsdom on Node.js 24.18.0, in development and production builds.

Additional context

Reproduced on next at efcaeba.

Possibly related to #3409/#3425, but this uses an ordinary signal update and a synchronous memo after the slow async memo. Replacing copy with a plain getter also removes the problem.

This reduced repro passes on #3337 at 20f3143 and its base b5bd6fb, so that passing result does not establish that #3337 fixes the current failure.

[ryansolid]

Fixed on next in a5d8eae via #3445 — thanks for the precise reproduction and the two controls; they made the trace straightforward.

Root cause: the isPending probe effect runs under its companions' optimistic lane, and when it pulled copy (a sync memo over the in-flight slow) to refresh it, recompute left that lane in force for the memo's run. Under a lane, a pending node on no lane serves its committed value instead of throwing NotReady, so copy published a stale "settled" value, dropped its pending status, and stopped holding the transaction — which then settled on fast's landing with slow still in flight.

Rule (A31 in the async spec): a memo computes under its own lane posture, never its puller's. recompute now clears the ambient lane for a non-effect node before re-establishing whatever lane the memo itself owns; effects keep the ambient lane. copy's pull throws NotReady as a plain reader would, and both reveal together.

Pinned by tests/ispending-combined-atomic-3442.test.ts (signals, manual clock) and test/ispending-combined-atomic-3442.spec.tsx (web, jsdom port of the playground), each with the plain-getter and no-probe controls.

— Claude via Cursor