fix(signals): a Loading on reset collects what its forwarded readers still wait on (#3459)

[brenelz]

Fixes #3459.

The gap

const fast = createMemo(async () => a());
const slow = createMemo(() => delay(2000, b()));
<p>B: {b()}</p>
<Loading on={a()} fallback={<p>Loading</p>}>
  <p>Fast: {fast()}</p>
  <p>Slow: {slow()}</p>
</Loading>

setB(1); await delay(500); setA(1). The first write is held by slow's flight (B stays 0, as it should). The second write changes on, the boundary resets, and content came straight back with B: 1 | Fast: 1 | Slow: 0 for the remaining 1.5 s, then Slow: 1. Expected: B: 1 | Loading until the b=1 answer lands, then Fast: 1 | Slow: 1 together.

Traced: the on reset (CollectionQueue.notify) clears _initialized and _sources, wakes the parked transaction, and then rebuilds _sources from the pending notifications that follow. Only the Fast effect notifies: it has a fresh flight for a=1. The Slow effect was already pending on slow from the earlier write, and status propagation dedupes on the reader's _pendingSources, so it never re-notifies. _sources ends up {fast}. The wake re-judges the b=1 transaction; by the #3375 ruling a reader behind a collecting boundary no longer blocks (reporterBlocksSource), so it commits and B: 1 shows. Then fast lands in a microtask, _checkSources empties, _disabled flips, and the tree reveals with slow still in the air. The hold left the lane, as ruled, but never arrived at the boundary. The #3375 fix collected the notifying effect's own _pendingSources, which is why one effect reading both fast() and slow() is fine; two sibling effects, the JSX shape, are not.

Fix

A reset moves the hold onto the boundary. The one place a forwarded reader is recorded is the live transactions' _asyncReporters (INV-3), so the reset walks them: for every reporter this boundary routes (_holds: its _queue chain reaches this queue with no collecting pending-type boundary in between, the same test reporterBlocksSource applies, plus the zombie/disposed filter), it adds the registered source and that reporter's _pendingSources, and flips to the fallback if it found any. wakeParked() still follows, so the transaction commits as before; what changes is that slow is now in _sources, and _checkSources keeps the fallback until its status clears.

scheduler.ts exports the transitions set for this; no scheduling logic changes. The scan is O(registered reporters) once per on change.

Trade-offs, stated plainly: the boundary now shows the fallback at the reset instant when it found a still-pending forwarded reader, before the new flight's own notification arrives. That is the ruled behaviour for a reset (new question, fallback until the answer). A reader pending only on a transaction-less re-ask (refresh()) has no reporter entry and is not collected; consistent with A9 (a re-ask is silent) and not the reported shape.

Verification

• @solidjs/signals: 180 files / 2037 tests green (1 skipped, pre-existing), including the new tests/loading-reset-collects-forwarded-3459.test.ts (the sibling-effect shape and the one-effect control, both pinned to B: 1 | Loading at the reset and Fast: 1 | Slow: 1 together).
• @solidjs/web: 88 files / 792 tests green, including the new test/loading-reset-collects-forwarded-3459.spec.tsx, a jsdom port of the issue's playground (real click, DOM sampled every 25 ms). Fails on next at t=50: B: 1 | Fast: 1 | Slow: 0, passes here.
• Core floor: 23,752 on next and 23,752 here, +0 B (boundaries.ts is outside the core-reachable set; the scheduler change is an export keyword).
• Brotli caps against a pristine next build: one cap ratcheted, app: CSR with Show/For/Loading/Errored/lazy 14.45 -> 14.5 KB, measured 14,467 against 14,429 (+38 B: the reset walk in boundaries.ts, retained wherever Loading is; the note in .size-limit.js cites it). The other Loading-carrying app scenarios grow 57 to 80 B and stay under their caps (hydrating + every store family lands at 28,848 against its 28.85 KB cap, 2 B of room). signals: + createStore is pre-existing red on next by 9 B (15,809 vs 15,800, from 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); here it measures 15,747, under the cap by brotli layout drift from the new scheduler export, not a real shrink, and its cap is left alone. Baseline is a forced (turbo --force) build of a8a8949 in a clean worktree; it matched the shared baseline byte for byte.
• Docs: INTERNALS-ASYNC-STATE.md extends the 2.0.0-rc Loading reveals stale content before downstream async work finishes #3375 bullet with the reset's collection rule; 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: 6308a0a

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 degrade performance by 5.8%

❌ 1 regressed benchmark
✅ 171 untouched benchmarks

Warning

Please fix the performance issues or acknowledge them on CodSpeed.

Performance Changes

	Benchmark	BASE	HEAD	Efficiency
❌	createStore setter: delete + set one root key (#3044 overlay)	549.4 µs	583.2 µs	-5.8%

Tip

Investigate this regression by commenting @codspeedbot fix this regression on this PR, or directly use the CodSpeed MCP with your agent.

---

_{Comparing brenelz:fix/loading-reset-collects-forwarded-3459 (6308a0a) with next (347a5ca)}