fix(signals): a direct-commit reader is neither entered nor born held — until() after a staged frame (#3482)

[ryansolid]

Closes #3482.

Bug

yield until(predicate) inside an action() never resolved when the predicate read the optimistic store a live stream feeds and the confirming frame landed before until() was called — the natural ordering for a server that broadcasts to subscribers before it answers the mutation. With a timeout, TimeoutError; without one the action hung forever. This is the single-primitive shape the until() docstring recommends.

Regression from d80cd1f6 (#3451, A29 creation-time born-held). 2.0.0-rc.8 is unaffected; next fails through c54e4c0ca.

Cause

A reader created mainline under a hold is born held (enterStagedRead sets stagedEntry; recompute stamps it, skips its synchronous first run and adds it to the transaction's _gatedSubs to replay at the commit). For a promise-delivery effect the commit is the action's settle — which its own promise is holding open. recompute's DIRECT_COMMIT arm exists for exactly this case ("deadlocks the hold (until)") but is gated on bornHeld === null, so the born-held path pre-empted it.

Fix

enterStagedRead's mainline arm exempts CONFIG_DIRECT_COMMIT readers — resolve(), until(), awaitable refresh()'s waiter — beside the existing verdict-pull exemption. They read staged truth by contract (the tunnel that keeps a hold deadlock-free) and apply on their own microtask, not the transaction's stashed queues, so they neither enter the transaction nor are born held. stagedEntry never gets set for them; every downstream bornHeld === null holds and they take the existing direct-commit path. The predicate sees the held frame, the action settles, the commit reveals the frame with the overlay revert.

-    if (GlobalQueue._verdictPull) return;
+    if (GlobalQueue._verdictPull || ctx._config & CONFIG_DIRECT_COMMIT) return;

Tests

tests/until-held-frame-before-call.test.ts — the issue's reproduction (frame lands while the action awaits its answer; until() reached after) and the control ordering (until() subscribed first, then the frame). On next the first fails with TimeoutError and the control passes; with the fix both pass.

signals 2,735 (+2) / 2 expected fails unchanged; solid 620; web 825 / 986 / 185 against the rebuilt dist. until, until-entanglement, born-held, resolve, refresh-await, tree-shake floor and rules index green.

Size

~+20 B minified on a core-retained read path; read +20..+40 B brotli across the scenarios (−30 on observe — layout). Two app caps (hydrating (no stores), CSR with …) ratcheted 0.05 KB with notes; the in-package tree-shake floor did not move.

Reproduction, bisect and analysis by @brenelz.

[changeset-bot]

🦋 Changeset detected

Latest commit: a2297c7

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

✅ 175 untouched benchmarks

---

_{Comparing fix/until-born-held-3482 (a2297c7) with next (c54e4c0)}

[ryansolid]

Closing unmerged — superseded by #3491. Not a bug: the until(...) in yield until(...) is evaluated in the await continuation, before the yield re-enters the transaction, so the effect is created mainline and born held as A29 specifies. A bare yield; after the await puts the call inside the transaction, where recompute's existing DIRECT_COMMIT arm — scoped to the effect's own held transition — already handles it; verified on next in #3491.

The exemption here was also too broad: CONFIG_DIRECT_COMMIT covers resolve() and awaitable refresh()'s waiter, which are routinely called outside actions. From mainline against a foreign hold they would have served that action's uncommitted frame instead of the committed view — a leak, not a fix. The gap it exposes (nothing pins that posture for the direct-commit readers) gets its own tests-only PR.

— Claude via Cursor