test: harden http2-request-never-settles against late session errors

[pacocartones]

This relates to...

The flaky test/http2-request-never-settles.js (added in #5603), currently failing intermittently on main and painting unrelated PRs red. Update 2026-08-09: root cause identified upstream — this is nodejs/node#64841, a V8 Maglev SIGSEGV regression in Node ≥ v24.15.0, not an undici bug. This PR no longer claims to fix the red CI; see the verification section.

Rationale

Since #5603 merged, this test fails intermittently in CI with a distinctive signature: ✖ test/http2-request-never-settles.js 'test failed' with no TAP output, no subtests, no stack, the file finishing in 3–9 s so the 15 s watchdog never runs. That signature is the test process dying abruptly, with the TAP stream discarded.

The original hypothesis of this PR was a late unhandled 'error' on the churn server (the only emitter without a handler). That hypothesis has now been falsified as the cause of the red CI (see below): the killer is a native crash, not a JS exception. The handlers remain correct hygiene for a churn server — an unhandled 'error' would kill the process (with a visible stack, unlike the flake) — but they do not cure the flake.

Changes

Test-only (+9/−0), copying the established pattern from test/http2-abort.js:

• server.on('error', …), session.on('error', …) and session.socket?.on('error', …) (tracked per session), and server.on('secureConnection', …) — late errors are swallowed instead of killing the process.
• A comment explaining why, so the next reader doesn't "clean them up".

No library code touched; the settle invariant under test is unchanged.

Verification

Red→green falsification: NEGATIVE — this hardening does not fix the flake. Reproduced locally on WSL2 (Ubuntu 24.04) using the exact recipe from nodejs/node#64841 (batches of 6 processes, taskset -c 0,1), on undici 86b62998:

Config	Result
v24.18.1, original test	1/72 and 4/72 SIGSEGV (exit 139) in two independent run sets
v24.18.1, test with this PR	1/72 SIGSEGV — same rate, same signature
v24.19.0 (current LTS), original test	4/72 SIGSEGV — the V8 backport in 24.19.0 does not cure it
v24.18.1 --no-maglev	0/48 SIGSEGV, but 1/48 SIGABRT (glibc corrupted size vs. prev_size) — not a clean stopgap
Controls: v24.14.1 (336 runs incl. ~30k amplified requests), v24.18.1 on Windows (72 runs)	0 failures — matches the issue being Linux-only, Node ≥ 24.15.0

The crash log is byte-identical to the CI signature: a single TAP line (✔ every h2 request settles under connection churn (seed 7)) and then sudden death — no error, no stack, exit 139. A probe test dying by an unhandled EventEmitter 'error' does show its stack under the test runner, confirming the CI killer is a signal death (SIGSEGV in Maglev concurrent compilation), not a JS exception.

What this means, honestly:

• The red on main and on fix(env-http-proxy-agent): ignore trailing dots when matching no_proxy #5637/fix(deduplicate): settle the deduplicated request when it is aborted #5673/fix(retry): honor the abort signal during the retry backoff #5675 is Intermittent SIGSEGV in Maglev concurrent compilation on Linux (PersistentHandles::GetHandle) — regression between v24.14.1 and v24.15.0 node#64841. Updating Node in CI does not help today (24.19.0 is still affected); --no-maglev is not a clean workaround (residual native instability).
• This PR's handlers defend a different, real hole (a late EPIPE/ECONNRESET on the churn server would crash the process with a stack), copied from the in-repo precedent. Merge it as hardening or close it — maintainer's call; it should not land under the expectation that it turns CI green.
• If undici wants green CI before the V8 fix lands, the only real lever is skipping/marking this file on Linux with Node ≥ 24.15.0 — a maintainers' decision, deliberately not proposed as a diff here.

Status

• I have read and agreed to the Developer's Certificate of Origin
• Tested
• Benchmarked (optional)
• Documented (test-only change)
• Review ready

[codecov-commenter]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 93.44%. Comparing base (dd85997) to head (d8de0c4).
⚠️ Report is 1 commits behind head on main.

Additional details and impacted files

@@           Coverage Diff           @@
##             main    #5674   +/-   ##
=======================================
  Coverage   93.43%   93.44%           
=======================================
  Files         110      110           
  Lines       38733    38776   +43     
=======================================
+ Hits        36190    36234   +44     
+ Misses       2543     2542    -1     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.