Fix flaky tcp_server specs under valgrind

[Watson1978]

Problem

test_run and test_run_timeout send data, sleep TIMEOUT (10ms), then tear the server down and assert on what arrived:

send_data(data) if data
sleep timeout          # <- assumes accept + on_read both happen within 10ms
server.detach
running = false
thread.join
@data                  # nil if they didn't

Under valgrind the reactor is an order of magnitude slower, so 10ms is often not enough and the example fails with expected: "hello", got: nil.

Seen on the Memcheck workflow, job 4.0 on ubuntu: https://github.com/socketry/cool.io/actions/runs/30615723672/job/91108332098

The other four matrix jobs (3.2, 3.3, head, and 3.4 before fail-fast cancelled it) passed on the same commit, and the same job passes on main, so this is timing, not a regression.

Fix

Wait for the data instead of for the clock, wrapped in Timeout.timeout so the wait can never turn into an indefinite sleep:

WAIT_TIMEOUT = 5.0

def wait_until(timeout = WAIT_TIMEOUT)
  Timeout.timeout(timeout) do
    sleep 0.001 until yield
  end
end

Five seconds is generous enough that a slow machine still passes, and a genuine hang now raises Timeout::Error at the wait itself instead of surfacing later as a confusing nil comparison. Both helpers also reset @data first, so the wait condition cannot be satisfied by a value left over from an earlier example.

Only these two helpers are affected. test_run_once already blocks on thread.join, and the two run_once(timeout) examples do not assert on @data.

Verification

Shrinking TIMEOUT to 0.0001 stands in for valgrind's slowdown and reproduces the CI failure locally:

before: 6 runs -> 4 failed (./spec/tcp_server_spec.rb:127)
after: 10 runs -> 0 failed

The timeout path was checked separately with a condition that never becomes true, which raises Timeout::Error at the deadline rather than hanging.

rspec: 58 examples, 1 failure (detach_race_condition_spec.rb:35) — a pre-existing Windows failure unrelated to this change.

🤖 Generated with Claude Code