go: 1.26.8, whose race detector leaves a forked child alone before exec on macOS (#1184)

macOS race CI failed now and then with "write |1: file already closed"
on a Claude subscription run's first prompt
(TestClaudeAuthFailureFallsBackWithoutRetryingTheLogin in 37605014029,
TestClaudeLastWordsAreRead in 37464400652; 37612795871 got the same
failure as "broken pipe"). The Claude Code stand-in had already ended
when the first prompt was written, and Wait had closed its stdin.

It never ran. On the fork's macOS runners, a debug build that recorded
the process's status at the failed write caught 8 cases in 10 jobs.
Each was "signal: segmentation fault", under 1ms of CPU, nothing on
stderr, with no magpie abort or group kill. The crash report
(gateway.race-*.ips) says "crashed on child side of fork pre-exec":

  __tsan::TraceSwitchPartImpl
  syscall.rawSyscall
  syscall.forkAndExecInChild
  syscall.forkExec
  os.StartProcess
  exec.(*Cmd).Start
  proc.StartTree
  gateway.(*subscriptionRun).launch

This is golang/go#79804. From Go 1.26.0, darwin's raw syscall wrappers
are instrumented under -race, and forkAndExecInChild calls them in the
child of fork. Go 1.26.5 marks them //go:norace (#79806); go.mod's
1.26.3, which setup-go builds CI with, has the bug. Only -race test
binaries on macOS are affected, so released builds never were. Any
os/exec start in a macOS race test could hit it; the Claude
subscription tests were where it surfaced. Local runs on go1.27.1
couldn't reproduce it, because that release already has the fix.

1.26.8 is the latest 1.26 patch release.

go.mod

@@ -1,6 +1,6 @@
 module github.com/yetone/magpie
 
-go 1.26.3
+go 1.26.8
 
 require (
 	github.com/Calcium-Ion/moejs v0.1.0-alpha.4