examples/napi/__tests__/issue-3119-repro.spec.ts (1)

9-12: Stale comments: these describe the pre-fix behavior, not the current expectation.

Lines 10-11 say "This test should fail/crash with the current implementation" — but with this PR's fix applied, the test should pass. Update the comments to reflect that this is a regression test.

📝 Suggested comment update

-// Exact reproduction of issue `#3119`
-// This test should fail/crash with the current implementation
-// Expected behavior: completes 5000 iterations
-// Actual behavior with bug: crashes at ~559 iterations with "malloc(): unaligned tcache chunk detected"
+// Regression test for issue `#3119`
+// Before the fix, this would crash at ~559 iterations with "malloc(): unaligned tcache chunk detected"
+// due to the instance being GC'd while the async generator was still in use.

examples/napi/__tests__/hidden-property.spec.ts (1)

38-49: Silent pass when [[InstanceRef]] descriptor is missing.

If the property doesn't exist (e.g., due to a regression in the Rust side), descriptor will be undefined and the entire block is skipped—the test passes without verifying anything. Consider failing explicitly when the descriptor is absent.

Suggested fix

   const descriptor = Object.getOwnPropertyDescriptor(
     iterator,
     '[[InstanceRef]]',
   )
-  if (descriptor) {
-    t.false(descriptor.enumerable, '[[InstanceRef]] should be non-enumerable')
-    t.false(descriptor.writable, '[[InstanceRef]] should be non-writable')
-    t.false(
-      descriptor.configurable,
-      '[[InstanceRef]] should be non-configurable',
-    )
-  }
+  t.truthy(descriptor, '[[InstanceRef]] property descriptor should exist')
+  t.false(descriptor!.enumerable, '[[InstanceRef]] should be non-enumerable')
+  t.false(descriptor!.writable, '[[InstanceRef]] should be non-writable')
+  t.false(
+    descriptor!.configurable,
+    '[[InstanceRef]] should be non-configurable',
+  )

examples/napi/__tests__/async-generator-gc.spec.ts (1)

5-6: Tests restricted to macOS ARM64 only.

This means the GC-interaction tests will be silently skipped on Linux CI (and Intel Macs). If the intent is to only run where GC behavior is reliably reproducible, a brief comment explaining why would help future contributors understand this isn't an oversight.

examples/napi/src/async_generator_repro.rs (1)

28-45: self.current increments unconditionally, even past max.

When current >= max, the method returns None (signaling done) but has already incremented current. For a test repro this is harmless—JS stops calling next() after done: true—but it's a small correctness nit if this is meant to be an example others copy.

Optional: increment only when yielding a value

   fn next(
     &mut self,
     _value: Option<Self::Next>,
   ) -> impl Future<Output = Result<Option<Self::Yield>>> + Send + 'static {
     let current = self.current;
     let max = self.max;
-    self.current += 1;
+    if current < max {
+      self.current += 1;
+    }
     async move {

.github/workflows/zig.yaml (1)

182-186: Mounting path differs from the other steps.

This step mounts to /napi-rs (-w /napi-rs) while the other two steps mount to /build (-w /build). This is likely intentional (carried over from the previous action config), but worth a quick sanity check that nothing depends on a consistent mount point across steps.