Keep a stub WEBrick from hanging the suite on Ruby 4.0

[tricknotes]

Problem

CI jobs on Ruby 4.0 intermittently stop producing output after an ember build and run until the 15-minute cap cancels them, e.g. run 33876168453 (build (4.0, 8.1, 7.0.0) and build (4.0, main, 7.0.0)), and before it runs 33875447042, 33871839705, 33869563469 and 33869331315. Every cancelled job in those runs is a Ruby 4.0 job; no Ruby 3.3 / 3.4 job ever hung. The --disable-dev-shm-usage / :js timeout from #668 did not stop it (the run above is the merge of #668).

Cause

Reproduced locally with Ruby 4.0.6 and a watchdog that dumps every thread's backtrace once an example runs for 90 seconds. The main thread was stuck here:

spec/lib/ember_cli/dev_server_spec.rb:246:in 'Thread#join'
spec/lib/ember_cli/dev_server_spec.rb:246:in 'NullServer#shutdown'
spec/lib/ember_cli/dev_server_spec.rb:258:in 'block (2 levels) in <top (required)>'   # after { servers.each(&:shutdown) }

while the WEBrick thread it joins sat in IO.select inside WEBrick::GenericServer#start with status == :Running.

NullServer#listen starts WEBrick with Thread.new { @server.start } and returns at once; the example's after hook then calls GenericServer#shutdown and joins the thread. GenericServer#shutdown only works once #start has set up its shutdown pipe (stop only flips :Running to :Shutdown, and alarm_shutdown_pipe does nothing while @shutdown_pipe is nil). When the thread has not run yet, shutdown is a no-op, the server boots afterwards and loops in IO.select forever, and the join never returns. On Ruby 4.0 a freshly created thread usually has not run by the time the example finishes.

Measured with a standalone loop of the same steps (boot WEBrick in a thread, connect once, shutdown, join with a 2s limit):

Ruby	hangs
4.0.6	187 / 300
3.3.6	0 / 300

The :js timeout from #668 cannot catch this: these are plain examples. And the progress dots of the examples that ran after the last ember build never reached the job log, because the Actions runner only flushes complete lines, which is why the jobs looked stuck right after the build output.

Fix

NullServer#listen now waits until WEBrick reports :Running (with a 5-second limit) before returning, so shutdown always reaches a server loop that can be stopped.

Verification

Ruby 4.0.6, Rails 8.1, Ember 7.0.0:

	before	after
rspec spec/lib/ember_cli/dev_server_spec.rb, 25 runs with a 25s limit	23 hung	0 hung
bin/rake (full suite)	—	148 examples, 0 failures

Also green on Ruby 3.3.6.

🤖 Generated with Claude Code

https://claude.ai/code/session_01FRn3BbAGhG1oWGbNkSHkvA

---

Generated by Claude Code