Fix flaky async_watcher_spec on macOS Ruby 3.2+

[Watson1978]

Problem

spec/async_watcher_spec.rb fails intermittently on macOS with Ruby 3.2+.
The handshake on line 33 uses rd.sysread(1) to wait for each child to become ready:

nr_fork.times { expect(rd.sysread(1)).to eq('.') }

Root cause (macOS Ruby 3.2+, platform-conditional)

On macOS Ruby 3.2+, IO.pipe returns pipes with O_NONBLOCK already set on both ends:

rd F_GETFL = 0x4  (= O_NONBLOCK on macOS)
wr F_GETFL = 0x5  (= O_WRONLY | O_NONBLOCK)

This is unrelated to cool.io or libev — a bare IO.pipe call with no other code loaded reproduces it. sysread calls read() directly without any EAGAIN retry, so if the child has not yet written '.' when the parent reads, Errno::EAGAIN is raised.

Reproduced without cool.io using a minimal script (100 iterations × 2 reads):

Out of 200 sysread calls: 50 EAGAIN  (platform: arm64-darwin24, ruby: 3.2.6)

Linux is unaffected because IO.pipe there does not set O_NONBLOCK.

Fix

Replace sysread with read, which handles EAGAIN internally and waits transparently for data regardless of the fd's nonblocking flag.

# before
nr_fork.times { expect(rd.sysread(1)).to eq('.') }

# after
nr_fork.times { expect(rd.read(1)).to eq('.') }

Assertions (lines.size == nr_signal, uniq.size == nr_fork) are unchanged.

Verification (macOS / Ruby 3.2.6)

• After fix: 0 failures out of 200 reads in the amplified race script
• bundle exec rspec spec/async_watcher_spec.rb passed 10 consecutive runs
• Full suite (bundle exec rspec spec/) passed: 44 examples, 0 failures

🤖 Generated with Claude Code