From f6bf2011a0922f4a78f4ac9f7e7be8d514835df1 Mon Sep 17 00:00:00 2001
From: Zoltan Kochan <z@kochan.io>
Date: Tue, 15 Sep 2026 10:07:23 +0200
Subject: [PATCH] fix(registry-mock): wait for the reaped registry child to
 leave the process table (#14917)

`ChildCleanup::assert_reaped` sampled the process table once, immediately
after the kill. On Windows a terminated process stays listed until its last
handle closes, so the PID can outlive the `wait()` that reaped it and the
assertion fires on a child that is already gone. This made
`successful_startup_retains_owner_and_reuse_does_not_stop_registry` flaky on
the Windows runner; it is the one test of the three that has no await between
dropping the instance and asserting.

Poll for up to ten seconds instead of sampling once. A child that is genuinely
left running still fails the assertion, verified by removing the kill from
`MockInstance::drop` and watching the test fail after the full budget.


Claude-Session: https://claude.ai/code/session_01GP1WscFutimgTWpePgVSV7

Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
---
 .../registry-mock/src/mock_instance/tests.rs  | 28 +++++++++++++------
 1 file changed, 19 insertions(+), 9 deletions(-)

diff --git a/pnpm/tasks/registry-mock/src/mock_instance/tests.rs b/pnpm/tasks/registry-mock/src/mock_instance/tests.rs
index 514e362e9f..71cc9549ad 100644
--- a/pnpm/tasks/registry-mock/src/mock_instance/tests.rs
+++ b/pnpm/tasks/registry-mock/src/mock_instance/tests.rs
@@ -6,7 +6,7 @@ use sysinfo::{Pid, ProcessRefreshKind, RefreshKind, Signal, System, UpdateKind};
 use tokio::{
     net::{TcpListener, TcpStream},
     task::JoinHandle,
-    time::{Duration, timeout},
+    time::{Duration, Instant, sleep, timeout},
 };
 
 fn options(client: &Client, port: u16) -> MockInstanceOptions<'_> {
@@ -28,6 +28,10 @@ fn processes() -> System {
     )
 }
 
+/// How long [`ChildCleanup::assert_reaped`] waits for the process table to
+/// catch up before it calls the child abandoned.
+const REAP_TIMEOUT: Duration = Duration::from_secs(10);
+
 /// Last-resort kill for the registry child a test spawned, so a failing
 /// assertion never leaves the process behind.
 struct ChildCleanup(Option<Pid>);
@@ -35,12 +39,18 @@ struct ChildCleanup(Option<Pid>);
 impl ChildCleanup {
     /// Disarms once the child is confirmed gone, because the OS is free to
     /// hand that PID to an unrelated process afterwards.
-    fn assert_reaped(mut self) {
+    ///
+    /// Windows keeps a terminated process listed until its last handle
+    /// closes, so the PID can outlive the `wait()` that reaped it. Polling
+    /// still fails a child that is genuinely left running, it just does not
+    /// race the process table.
+    async fn assert_reaped(mut self) {
         let pid = self.0.expect("cleanup is armed");
-        assert!(
-            processes().process(pid).is_none(),
-            "registry child {pid} must be stopped and reaped",
-        );
+        let deadline = Instant::now() + REAP_TIMEOUT;
+        while processes().process(pid).is_some() {
+            assert!(Instant::now() < deadline, "registry child {pid} must be stopped and reaped");
+            sleep(Duration::from_millis(50)).await;
+        }
         self.0 = None;
     }
 }
@@ -90,7 +100,7 @@ async fn cancelled_startup_reaps_child() {
     startup.abort();
     let error = startup.await.unwrap_err();
     assert!(error.is_cancelled(), "aborting startup cancels it: {error}");
-    child.assert_reaped();
+    child.assert_reaped().await;
 }
 
 #[tokio::test]
@@ -103,7 +113,7 @@ async fn failed_readiness_check_reaps_child() {
         .expect("closed connection must fail the readiness check")
         .unwrap_err();
     assert!(error.is_panic(), "readiness failures panic: {error}");
-    child.assert_reaped();
+    child.assert_reaped().await;
 }
 
 #[tokio::test]
@@ -118,5 +128,5 @@ async fn successful_startup_retains_owner_and_reuse_does_not_stop_registry() {
     assert!(registry.spawn_if_necessary().await.is_none(), "a ready registry is not respawned");
     assert!(registry.is_registry_ready().await, "reuse must leave the registry serving");
     drop(instance);
-    child.assert_reaped();
+    child.assert_reaped().await;
 }