Node 24 / V8 13 compatibility with CORO_PTHREAD coroutines

[jbaczuk-qualia]

Summary

Makes fibers build and run on node 24 (V8 13.6) while keeping the CORO_PTHREAD coroutine backend that binding.gyp on async-resource already selects for every Linux build (and that arm64 production compiles today). Five commits on top of async-resource, nothing else:

Commit	What	Backend
33e592c compile against V8 13 and return the fiber's result from run()	SetAccessor -> SetNativeDataProperty, Holder() -> This(), kFinalizer -> kParameter weak callbacks; uni::Return() takes PropertyCallbackInfo by const reference (V8 13 stores the argument slots inline, so the by-value copy silently dropped every getter result); fibers_async.js returns the wrapped function's value from runInAsyncScope	any
5ecb0ab reset the handle when a suspended fiber is garbage-collected on V8 >= 10.4	kParameter is a phantom callback and V8 CHECKs that the handle was reset (Handle not reset in first callback); a yielded fiber whose JS object became unreachable aborted node 20+. Same change as #4, rebased	any
d8ef8e2 forward Fiber.poolSize from the async wrapper to the native setter	on V8 13, Fiber.poolSize = n on the wrapper creates an own property and the native pool size silently stays 120 (on node 18 the assignment reached the native setter). @qualia/patches relies on Fiber.poolSize = 1e9, so without this every finished coroutine beyond 120 concurrent ones was destroyed	any (found on node 24)
d3260f6 destroy the coroutine context before freeing its stack	~Coroutine freed the stack and then called coro_destroy(). With pthread coroutines the thread's struct pthread lives at the top of that user-supplied stack, so pthread_cancel() dereferenced unmapped memory: SIGSEGV on every coroutine destroy. This is why test/pool.js and test/cleanup.js have segfaulted on every pthread build, node 18 included	pthread (harmless elsewhere)
213c42d create bin/ recursively when installing from a source-only package	build.js only created bin/<platform>/; published tarballs carry bin/ because of the prebuilts, npm pack of a source tree does not, so npm ci compiled fibers and then failed the final rename with ENOENT	any

Not in this PR, on purpose (compare #2, the CORO_UCONTEXT port, now closed): no backend switch, no V8 ThreadId TLS key discovery, no v8_qualia_set_thread_stack_start hook. With pthread coroutines every fiber is a real OS thread, so V8's own per-thread state, its Locker/Unlocker thread archiving and cppgc's stack scan all work as designed and need no help from fibers or from V8 patches.

Node requirements

Runs on stock node 24 except for process.binding('async_wrap'), which stock node 24 refuses: setupAsyncHacks then disables itself and every yield corrupts the async hook stack (async hook stack has become corrupted, 7 of 19 suite failures on stock 24.21.0). qualialabs/node PR for custom-v24-pthread restores the allowlist entry and fixes node 24's thread_local TLS root cert store, which otherwise aborts on the first HTTPS request from a fiber thread. The Meteor server must run with --no-async-context-frame (Meteor's executionAsyncId() === 0 guard).

Verification

All on arm64 Linux (M-series container), node = qualialabs/node custom-v24-pthread (24.21.0 + 3 node-only patches, no V8 changes) unless stated.

Test	Result
test/*.js (20 files incl. orphan-gc.js)	20/20. async-resource as shipped: 16/19 on node 18 pthread (pool.js, cleanup.js, future-exception.js)
pool.js / cleanup.js / orphan-gc.js before d3260f6	SIGSEGV in __pthread_cancel <- coro_destroy <- Coroutine::~Coroutine <- Coroutine::run (gdb, -g -O0 build)
d3260f6 applied to fibers 5.0.4 on node 18.20 (pthread)	pool.js, cleanup.js and a 5-concurrent-coroutine destroy test pass; unpatched 5.0.4 segfaults on all three
GC stress: 8 fibers x 20 yields per round, heavy allocation, AsyncLocalStorage check, timer/immediate/nextTick resumes, --stress-incremental-marking --stress-compaction	100 rounds clean with AsyncContextFrame on, 100 with --no-async-context-frame, 300 with --stress-scavenge=50, 50 with poolSize = 4 (coroutines destroyed every round)
8 fibers each doing an HTTPS round trip that yields until the response, plus tls.createSecureContext({ca}) from 3 fiber threads	pass (aborts in CleanupQueue::Add on stock node 24)
global-deployment-center in qli5 on a Graviton2 remote-dev host	boots, HTTP 200, DDP login 403 + subscription ready, no failure signatures, 180 OS threads for the pooled fibers (see the comment below and qualialabs/qualia#56473)

Publishing

package.json still says 5.0.5, which is already published. Publish as 5.1.0 and pin it explicitly in node 24 services; ^5.0.4 consumers on node 18 keep 5.0.4 untouched. d8ef8e2 and d3260f6 also apply to node 18 and could go out as a 5.0.x later after a canary, separately from the node 24 work.

🤖 Generated with Claude Code

[jbaczuk-qualia]

GDC on the Graviton2 remote-dev host (2026-09-21, this branch + qualialabs/node#6 custom-v24-pthread binary, qli5, 4 cores): boots and serves (HTTP 200, Server startup complete, sync.data ran at boot), DDP login -> 403, meteor.loginServiceConfiguration subscription -> ready (2 runs), 26 yields observed in fiber_yield_switch_seconds, no Meteor code must always run within a Fiber, CleanupQueue::Add, async-hook-stack, Handle not reset or segfault lines in docker logs, 0 restarts after boot. Server child runs with --no-async-context-frame, 180 OS threads (one per pooled fiber), RSS 475 MB idle. Still to do: browser click-through, Lucky, --watch restarts, load. Details in gist da31bb93ceb6f57a66f1b47c1415dee7.