Bug: process does not exit on Ctrl+C — `serve()` polling loop keeps process alive with no signal handler

[momics]

Bug

When node.serve() is called, the rawServe() polling loop keeps an unresolved Promise alive. This prevents the Deno (and Node.js) process from exiting naturally, even after all user code has finished.

There is no default SIGINT handler registered, so pressing Ctrl+C does not stop the process.

Reproduction

const node = await createNode();
node.serve((req) => new Response("hello"));
// process hangs forever — Ctrl+C does nothing

The only ways to exit are:

• Pass an AbortSignal to serve({ signal }) and abort it manually
• Call node.close() explicitly

Neither is documented as required.

Expected behaviour

At minimum, the documentation and/or the getting-started example should show that node.close() / an AbortSignal is required for the process to exit.

Ideally, iroh-http should register a default SIGINT/SIGTERM handler (similar to how Deno.serve() handles this) so that Ctrl+C gracefully shuts down the serve loop and exits. This could be opt-out:

node.serve({ handleSignals: true }, (req) => new Response("hello"));

Affected platforms

• Deno (verified)
• Node.js (likely same issue)

[momics]

Additional note: the likely root cause is that the rawServe() polling loop never resolves its Promise — it blocks waiting for the next request indefinitely. The process isn't just missing a signal handler; the serve loop itself is in a permanent suspended state. Until stopServe() is called (which drains the loop and lets the Promise resolve), the JS runtime has no way to know it should exit. A SIGINT handler would fix the symptom, but the core issue is that the serve loop provides no natural exit path without an explicit shutdown call.

[momics]

Fixed in 8fc0a9b — fix(shared): await loopDone before resolving finished (#115).

Root cause: finished = Promise.race([loopDone, onNodeClose]) let finished resolve via onNodeClose before the pending nextRequest FFI op completed, leaving a dangling async op that kept the process alive. Fix chains onNodeClose on loopDone so finished only resolves once the polling loop has actually exited. The regression test now runs with sanitizeOps: true (no ignore).

[momics]

Reopening — the regression test (serve — no pending ops remain after signal abort) still times out in CI (run 25101906065).

The ignore: true flag was accidentally removed in 8df1e97 (the #122 fix), while the comment still said "it is marked ignore". Re-adding ignore: true in the next commit so CI stays green, but the underlying issue (pending FFI op after stopServe) is not resolved.

[momics]

Fixed in bd64fa2.

Root cause (re-emergence): #122 changed polling from async FFI to sync poll + MessageChannel yield. When stopServe fires while the loop is in await yielder.yield(), the loop can't break until the next MessageChannel tick fires AND the subsequent sync poll returns the -1 sentinel. Under CI load, the stopServe FFI call hasn't completed by the time the loop re-polls, causing a timing-dependent timeout. The leaked MessageChannel ports also trigger Deno's resource sanitizer.

Fix: createYieldFn() now returns a cancel() callback that immediately resolves any in-flight yield. Both stopServe callsites invoke the canceller via a per-endpoint serveCancellers Map before dispatching the FFI stop call, so the loop re-polls promptly and sees the -1 sentinel.

Test is un-skipped and passes with sanitizeOps: true.