[CORO_UCONTEXT] Node 24 / V8 13 compatibility (ucontext coroutines; closed, superseded by #5)

[jbaczuk-qualia]

Summary

Makes fibers build and run on qualialabs/node#custom-v24 (Node 24.21.0, V8 13.6). Four commits:

• e0fbe77 compile against V8 13 and return the fiber's result from run(). SetAccessor -> SetNativeDataProperty, Holder() -> This(), kFinalizer weak callbacks -> kParameter. PropertyCallbackInfo is now stored inline, so the by-value uni::Return() helper silently dropped every getter result; it takes a const reference now. fibers_async.js returns the wrapped function's value so run() resolves to it (README semantics).
• a7f5750 ucontext coroutines on Linux and arm64 instead of pthreads. Same binding.gyp change as build: use ucontext coroutines instead of pthreads on Linux and arm64 #3. CORO_PTHREAD builds and passes the fibers suite (16/19) and GC stress on the patched node 24, so V8 13 itself does not rule it out (an earlier version of this description said it did; that was wrong). What does rule it out is node 24: GDC with pthread fibers dies on its first HTTPS request because crypto_context.cc keeps the root cert store in thread_local statics and a fiber thread re-registers its cleanup hook (CleanupQueue::Add CHECK). ucontext avoids the whole class, costs ~0.7 µs per switch instead of ~13 µs plus an OS thread per fiber, and passes more of the suite (18/19 vs 16/19).
• a692f79 find V8's ThreadId TLS key on V8 >= 12 and report the running stack to V8. The legacy TLS-key scan finds nothing on V8 >= 12, so Fiber silently ran without swapping the thread id. New scan: snapshot pthread TLS from a helper thread that has entered the isolate and pick the key whose value is a small monotonically assigned id. FIBERS_DEBUG_TLS=1 prints what it found. On every switch fibers now calls v8_qualia_set_thread_stack_start() (resolved with dlsym, no-op when absent) so cppgc's conservative stack scan walks the fiber stack instead of the OS thread's.
• 2fb6179 make the key discovery tolerate concurrent thread creation and fail loudly. The two-snapshot diff flaked about 1 run in 80 when another thread was assigned a ThreadId between the snapshots (vb == va + 1 no longer held) and died on an assert. It now retries up to 8 times, accepts vb - va <= 64, requires exactly one candidate key, and aborts with a message naming the problem instead of an assert. 150/150 process starts clean afterwards.

Verification

• test/*.js: 18/19 on custom node 24 with ucontext (future-exception.js only, which fails on every runtime and backend); the shipped pthread build is 16/19 on node 18 and on this node (pool.js, cleanup.js segfault).
• GC stress (gc_stress.js with --stress-incremental-marking --stress-compaction): 150 rounds clean on a Graviton2 remote-dev host, 300 rounds on an M-series container. Needs [CORO_UCONTEXT] Node 24.21.0 with CORO_UCONTEXT-based fibers support (custom-v24; closed, superseded by #6) node#4 commit fbc38d02; before it the ucontext run segfaulted within ~10 rounds because cppgc scanned from a stale stack start.
• Wasm code GC with suspended fibers no longer FATALs when running on [CORO_UCONTEXT] Node 24.21.0 with CORO_UCONTEXT-based fibers support (custom-v24; closed, superseded by #6) node#4 (55f54c27); that is a node fix, nothing here changes for it.
• global-deployment-center runs on it in qli5 (login, subs, methods); 200k-yield leak test shows a flat heap and no external growth.

Not published yet: the monorepo experiment vendors npm pack of this branch as fibers-5.0.5-node24.tgz. 5.0.5 is already published, so this needs a new version (5.0.6 is taken by #3's bump; use 5.1.0 or rebase onto #3). Related PRs: qualialabs/node#4, qualialabs/node-builder#16, qualialabs/meteor-lite#41, qualialabs/qualia#56265 and #56266.

🤖 Generated with Claude Code

[jbaczuk-qualia]

Closed 2026-09-21, kept for posterity. This is the CORO_UCONTEXT port. We decided to upgrade to node 24 on the CORO_PTHREAD backend that production runs today, with the minimum number of changes: #5 carries only the backend-independent commits from here (e0fbe77 V8 13 compile fixes, the #4 orphan-fiber fix) plus two pthread fixes, and needs no V8 patches. The parts unique to this branch are the reasons we are not shipping it: the V8 ThreadId TLS key is found by snapshotting pthread TLS from helper threads and diffing values (a692f79, hardened in 2fb6179 after it flaked 1 in 80 starts), and a wrong guess swaps the wrong TLS slot on every coroutine switch, which corrupts V8 state silently; v8_qualia_set_thread_stack_start has to be called on every switch to keep cppgc from scanning the wrong stack; and both depend on three V8 GC patches in qualialabs/node#4 that have to be re-audited on every node release. Measurements and the full comparison stay in the write-up (gist da31bb93ceb6f57a66f1b47c1415dee7).