fix: reset the handle when a suspended fiber is garbage-collected on V8 >= 10.4

[jbaczuk-qualia]

Problem

f3cbba6 ("Make compatible with v20", PR #1) registers the Fiber weak callback with WeakCallbackType::kParameter because V8 removed kFinalizer. The two have different contracts:

• kFinalizer (node 18, V8 10.2): the callback runs before the object is reclaimed and may resurrect it. Fiber::WeakCallback relied on this for suspended fibers: ClearWeak(), unwind later in DestroyOrphans(), MakeWeak() again.
• kParameter (node 20+): a phantom callback. The object is already gone and V8 CHECKs that the callback reset its handle (global-handles.cc: "Handle not reset in first callback").

So on node 20+ a yielded fiber whose JS object becomes unreachable (its resuming callback was dropped: an abandoned request, a cleared timer, a socket closed without firing) aborts the process at the next GC:

# Fatal error in , line 0
# Check failed: Handle not reset in first callback. See comments on |v8::WeakCallbackInfo|.

Finished and never-started fibers were fine (delete → destructor resets the handle), which is why the suite and the GDC click-throughs never hit it. Node 18 is unaffected.

Fix

On V8_AT_LEAST(10, 4):

• WeakCallback: for a started (yielded) fiber, queue it as an orphan and reset the handle instead of ClearWeak().
• DestroyOrphans: after UnwindStack(), delete the fiber instead of MakeWeak() on a handle that no longer exists.
• MakeWeak(), ClearWeak() and the Fiber.current getter return early on an empty handle. Fiber::Yield_ and user code in the zombie's catch/finally reach these while the stack unwinds; the first cut without the guards segfaulted in GlobalHandles::ClearWeakness.

The node 18 code path is unchanged.

Test

test/orphan-gc.js: 200 yielded fibers dropped, two forced GCs, a Fiber.run() to trigger DestroyOrphans, and a check that all 200 unwound (each touches Fiber.current from its catch).

build	result
node24-compat unpatched, node 24.21.0 (patched custom-v24, ucontext)	abort, Check failed: Handle not reset in first callback
this branch, node 24.21.0	pass; full suite 20/20
fibers 5.0.4 on node 18.16.1 (kFinalizer path)	pass

Notes

• Base is node24-compat (PR [CORO_UCONTEXT] Node 24 / V8 13 compatibility (ucontext coroutines; closed, superseded by #5) #2), so this lands with the V8 13 port. PR Async resource #1 carries the same kParameter change and should not merge without this.
• Found while checking the "node 20 removed the API that cleans up memory" concern: the removed API is kFinalizer; cleanup still runs, the regression is this crash.

🤖 Generated with Claude Code

[jbaczuk-qualia]

Closed 2026-09-21: carried into #5 as 5ecb0ab (rebased onto async-resource without the ucontext commits). This fix is backend-independent; it was based on node24-compat, the CORO_UCONTEXT branch that we are not shipping. With pthread coroutines the same test (test/orphan-gc.js) also needed the destructor-order fix in #5 (d3260f6): unwinding orphans beyond the pool size destroyed coroutines, and coro_destroy segfaulted on pthread.