From c6e357156fb7fe07093cc94a277a306a97db0b9e Mon Sep 17 00:00:00 2001
From: robobun <117481402+robobun@users.noreply.github.com>
Date: Mon, 29 Jun 2026 05:49:48 +0000
Subject: [PATCH 1/2] Quarantine test-net-connect-memleak on linux-x64-musl

test/js/node/test/parallel/test-net-connect-memleak.js fails on ~half of
PR builds on the alpine 3.23 x64 and x64-baseline lanes. It asserts a
FinalizationRegistry cleanup callback fires within a single globalThis.gc()
plus one setImmediate after the connect callback's closure is unreferenced.
The spec gives no timing guarantee for cleanup callbacks and JSC schedules
them via DeferredWorkTimer with no defined ordering relative to the immediate
queue, so on musl x64 delivery sometimes slips past that single setImmediate.

Its tls twin (test-tls-connect-memleak.js) has the identical onGC + gc() +
setImmediate pattern and was already quarantined on this exact matrix. Both
are verbatim upstream node ports we do not edit, so mark the net twin FLAKY
next to the tls one. It still runs on glibc Linux, darwin, and Windows.
---
 test/expectations.txt | 21 ++++++++++++---------
 1 file changed, 12 insertions(+), 9 deletions(-)

diff --git a/test/expectations.txt b/test/expectations.txt
index 24e09b7de2ad..179d9529c00b 100644
--- a/test/expectations.txt
+++ b/test/expectations.txt
@@ -12,20 +12,23 @@ test/cli/run/run-crash-handler.test.ts [ FAIL ] # automatic crash reporter > seg
 # the spawned child needs process.binding('inspector'), which is not implemented
 test/js/node/test/parallel/test-inspector-enabled.js [ FAIL ]
 
-# Verbatim node v26.3.0 test asserting a FinalizationRegistry callback fires
+# Verbatim node v26.3.0 tests asserting a FinalizationRegistry callback fires
 # within ONE globalThis.gc() + ONE setImmediate after the connect callback's
 # closure is unreferenced. The FinalizationRegistry spec gives no timing
 # guarantee for cleanup callbacks; JSC schedules them via DeferredWorkTimer
 # with no defined ordering relative to the immediate queue. The connect
-# listener IS removed (verified: listenerCount("secureConnect") === 0 in
-# done()) and the object IS collected (test passes 70/70 on darwin and
-# glibc Linux); on alpine x64 the FR callback delivery slips past the single
-# setImmediate after this PR's added module loads at process startup shift
-# the heap layout. The robust fix is gcUntil() rather than a single tick,
-# but the file is a verbatim upstream port. Quarantined on the failing
-# linux-x64-musl matrix only; still runs everywhere else (build 63145:
-# alpine 3.23 x64 + x64-baseline only).
+# listener IS removed (verified listenerCount 0 in done(): "secureConnect"
+# for tls, "connect" for net) and the object IS collected (both pass on
+# darwin and glibc Linux); on alpine x64 the FR callback delivery slips past
+# the single setImmediate, so the assertion reds the lane on ~half of PR
+# builds. The robust fix is gcUntil() rather than a single tick, but these
+# are verbatim upstream ports we do not edit. Quarantined on the failing
+# linux-x64-musl matrix only (alpine 3.23 x64 + x64-baseline); still runs
+# everywhere else. Both twins share the identical onGC + gc() + setImmediate
+# pattern: the tls one was quarantined first (build 63145), the net one has
+# been failing intermittently since ~2026-06-28 (e.g. builds 66653, 66657).
 [ LINUX-X64-MUSL ] test/js/node/test/parallel/test-tls-connect-memleak.js [ FLAKY ] # JSC FinalizationRegistry callback delivery vs setImmediate timing on musl x64
+[ LINUX-X64-MUSL ] test/js/node/test/parallel/test-net-connect-memleak.js [ FLAKY ] # JSC FinalizationRegistry callback delivery vs setImmediate timing on musl x64
 
 # Vendored node v26.3.0 stream tests blocked on missing native subsystems (see PR #31826)
 test/js/node/test/parallel/test-stream-pipeline.js [ SKIP ] # block at L271 hangs: pipeline(rs, req) writes 11x'hello' raw after a never-ended GET's \r\n\r\n; node's llhttp rejects lowercase 'h' as a method char (HPE_INVALID_METHOD -> clientError -> 400+close -> req 'close' -> pipeline callback fires), but bun's uWS HttpParser buffers any incomplete run of valid tchars waiting for the request-line, so the connection stays open and the callback never fires. Pre-existing server-parser leniency; needs uWS HttpParser to reject non-uppercase method bytes like llhttp.

From 3e810cfa179f9bf5fd6c2c1f76b432fc1a40ce17 Mon Sep 17 00:00:00 2001
From: robobun <117481402+robobun@users.noreply.github.com>
Date: Mon, 29 Jun 2026 07:07:50 +0000
Subject: [PATCH 2/2] ci: retrigger