fix(signals): an async memo's held landing keeps the committed frame's dependencies (#3461)

[brenelz]

Fixes #3461.

The gap

const [a, setA] = createSignal(0);
const [b, setB] = createSignal(0);
const slow = createMemo(() => delay(2000, b()));
const selected = createMemo(async () => (b() ? b() : a()));
// click: setB(1); await delay(500); setA(1);
<p>A: {a()}</p> <p>B: {b()}</p> <p>Slow: {slow()}</p> <p>Selected: {selected()}</p>

setB(1) is held by slow. Half a second later setA(1) publishes alone: A: 1 | B: 0 | Slow: 0 | Selected: 0 for the rest of the hold, and Selected no longer agrees with the inputs on screen (b() is 0, so it should read a()). At 2000 everything settles to 1. With selected a plain createMemo the a write is held with the transaction and the four values reveal as one frame.

Traced: selected's held pass reads only b (throws NotReady at the flight's registration, keeping its full dependency list with _depsTail at b). Its flight lands in a microtask into the hold, and asyncWrite called trimStaleDeps unconditionally, before the write, so the selected -> a edge was gone even though setSignal then staged the landed value under the transaction. The committed Selected: 0 still derived from a, but the mainline a write no longer reached selected, never entered its hold through the stamp, and committed beside the stale derivation. A30 (#3410) already rules this for a sync pass (recompute trims only when the pass published, commitPendingNode trims a staged pass) and for effects (#3438); the async landing predates the rule (the trim is from the initial import) and was the missing arm.

Fix

The landing follows A30's gate: asyncWrite trims after the write and only when the landing published (_pendingValue === NOT_PENDING, an equal-value or lane landing); a transition-held landing leaves the tail for commitPendingNode, exactly as a staged sync pass does. One call moved:

if (el._pendingValue === NOT_PENDING) {
  el._loading = false;
  if (wasReask) el._x!._reask = false;
  trimStaleDeps(el);
}

Now the a write reaches selected, its stamped recompute enters the hold, and the frame is 5000: A: 1 | B: 1 | Selected: 1 | Slow: 1, identical to the sync control.

Trade-offs, stated plainly: a plain (unheld) landing's trim moves from the landing to the same flush's commit, A30's stated cost model. In the #3373 shape (a landing held while another input is still pending, _error set) the stale tail lives until the next clean pass or landing, which is over-subscription only, never a missed wake.

Side finding, not fixed here: the equal-value variant selected = async () => (b() ? 0 : a()) still tears the same way in both the sync and the async shape, because A30 lets a pass that changed nothing trim at its tail (a landing equal to the committed value published nothing new). This PR mirrors that carve-out deliberately; closing it is a rule decision and is noted in the A30 paragraph.

Verification

• @solidjs/signals: 180 files / 2038 tests green (2036 passed, 1 skipped, plus the new file), including the new tests/async-landing-deps-3461.test.ts (the issue shape on a manual clock pinned to exact frames, and the sync control). The only non-green run was heap-mark-incremental "linear in N (2.0.0-rc.7: mounting N rows is O(N²) when each row creates an effect over a signal written during its own creation — markHeap re-walks the whole pure heap on every memo pull #3350)", a timing test that passed on re-run in isolation.
• @solidjs/web: 88 files / 793 tests green, including the new test/async-landing-deps-3461.spec.tsx, a jsdom port of the issue's playground (real click, DOM sampled through the hold, with the plain-memo control); fails on next, passes here.
• Core floor 23,752 -> 23,752 (0 B, a moved call). Brotli: the "signals: core floor" cap ratcheted 8.85 -> 8.90 KB, measured at 8851 B against 8849 on a pristine next build (+2 B, noise for the moved call); every other scenario is within noise of the baseline. "signals: + createStore" is pre-existing red on next by 9 B (15,809 vs the 15,800 cap, from the store changes in perf(signals,web,universal,html): merge/omit are always lazy views; consumers read the leaves #3454/test(signals): store oracle + four signal states; latest() seed leak and body-end visibility fixes #3455); this PR does not touch that cap and measures 15,762 on its own build.
• Spec: A30 gains the landing arm (mechanism line and a short paragraph, pinned by the new test); rules index regenerated (--check clean; the one unresolved test id is pre-existing on next). Changeset included.

🤖 Generated with Claude Code

[changeset-bot]

🦋 Changeset detected

Latest commit: c24e38e

The changes in this PR will be included in the next version bump.

This PR includes changesets to release 11 packages

Name	Type
@solidjs/signals	Patch
test-integration	Patch
@solidjs/web	Patch
@solidjs/babel-plugin	Patch
@solidjs/compiler	Patch
@solidjs/diagnostics	Patch
@solidjs/element	Patch
@solidjs/h	Patch
@solidjs/html	Patch
solid-js	Patch
@solidjs/universal	Patch

Not sure what this means? Click here to learn what changesets are.

Click here if you're a maintainer who wants to add another changeset to this PR

[codspeed]

Merging this PR will not alter performance

✅ 172 untouched benchmarks

---

_{Comparing brenelz:fix/async-landing-deps-3461 (c24e38e) with next (a8a8949)}