2.0.0-beta.19: Optimistic lanes: a shared subscriber merges lanes across transactions

[brenelz]

Version: 2.0.0-beta.19, branch next (reproduced @ c7bb2c8, post-#2899; the mechanism is untouched by #2902–#2904)

Summary

When one effect subscribes to keys/signals touched by two different actions, insertSubs → assignOrMergeLane merges the earlier action's lane into the later action's lane without merging their transactions. From then on resolveTransition reports the later action's transaction for the earlier action's nodes. Two consequences:

• same-key follow-up writes entangle with the wrong transaction (core/optimistic.ts:85 initTransition(resolveTransition(el))), and
• the 2.0.0-beta.19: createOptimisticStore: first-settling action wipes other in-flight actions' overrides #2899 owner-stamp/scoped-clear path consumes the earlier action's keys when the later action settles — reverting overrides that belong to a still-live transaction. This is the exact bug class 2.0.0-beta.19: createOptimisticStore: first-settling action wipes other in-flight actions' overrides #2899 was fixing, resurfacing through the lane channel.

Notably, the shipped #2899 test 3 passes only because of write order: mirror it with action B's two writes swapped (disjoint key first, same key second) and it fails.

Root cause

packages/solid-signals/src/core/lanes.ts:159 — mergeLanes(sourceRoot, existingRoot) inside assignOrMergeLane, reached from insertSubs (core/scheduler.ts:621), merges lanes from different transactions. resolveTransition (lanes.ts:118-123) then prefers the hijacked lane's _transition. Consumers: core/optimistic.ts:85, the store owner stamp (store/store.ts:621-624), and the scoped clear (store/optimistic.ts:155-193).

Repros

The shipped #2899 test with B's writes swapped

const [s, setS] = createOptimisticStore({ a: 1, b: 2 });
createRoot(() => createRenderEffect(() => s.a + s.b, () => {})); // ONE effect reads both
flush();

const pA = action(function* () { setS(d => { d.a = 10; }); yield gA.promise; })();
flush();
const pB = action(function* () {
  setS(d => { d.b = 20; });  // disjoint key FIRST — hijacks A's lane via the shared effect
  setS(d => { d.a = 11; });  // same key as A — should merge the transactions
  yield gB.promise;
})();
flush();
// [s.a, s.b] === [11, 20] ✓

gB.resolve(); await pB; flush();
[s.a, s.b]; // FAILS: [1, 2] — both keys reverted while A is still in flight.
            // Expected [11, 20]: merged same-key semantics hold until the LAST action.

Plain createOptimistic signals, three actions

One effect reads x() + y(). Action A writes x, action B writes y (hijacks x's lane into B's transaction), action C rewrites x — C entangles with B instead of A. When A settles, x reverts to committed while C is still in flight (observed 1, expected 11). A two-store variant fails the same way.

Impact

HIGH — one component reading two independently-mutated values is the normal case of optimistic UI, and overlapping actions silently revert each other's live overrides. The fix likely needs a design decision at the lane/transaction ownership boundary: either merge transactions when lanes merge, or key resolveTransition off the node's own transaction rather than the lane's.

[ryansolid]

Fixed in 5ab7d17 on next.

The diagnosis was right: lanes are scheduling affinity and the shared-subscriber merge is correct for effect flushing, but resolveTransition following the merged lane root's _transition let the lane channel transfer transaction ownership. Neither of the two proposed directions merged transactions on lane merge (that would over-entangle disjoint actions and undo #2899); instead ownership now lives on the node: every optimistic write stamps _overrideOwner (post-merge, so genuinely entangled same-key writers share the joint root; cleared at settle), and resolveTransition prefers a live owner stamp over the lane, falling back to the lane's transition for nodes without overrides (async routing) as before. This is the node-level sibling of #2899's STORE_OPTIMISTIC_OWNERS stamps, recorded as the A18 node corollary in the spec.

Both repros are pinned in tests/optimistic-lane-transaction-ownership.test.ts — the #2899-test-3 shape with B's writes swapped (disjoint first, same-key second) and the three-action plain-signal hijack. +118 B min / +34 B gzip; hot paths untouched.