v2.336.0 on macOS (arm64): Runner.Worker deadlocks in Process.Start — forked child never execs, spins at 100% CPU; every job hangs seconds after start #4570

Describe the bug

After the self-hosted runner auto-updated 2.335.1 → 2.336.0 (2026-07-22), every job on our macOS runner hangs a few seconds after starting, during an early process spawn (checkout/bash steps). 6 out of 6 jobs since the update wedged; all jobs before the update passed. Rolling back to 2.335.1 (same machine, same OS) fixed it immediately.

Mechanism, observed live on a wedged job:

• _diag/Worker_*.log ends with a Starting process: block that is never followed by
  Process started with process id N — the worker is stuck inside Process.Start.
• ps shows a second Runner.Worker spawnclient ... process — a forked child that never
  reached exec — burning ~100% of a core (physical footprint ~2 MB), child of the real Worker.
• sample (macOS profiler) on that child shows a single thread spinning in managed CoreCLR
  code (ThreadNative::KickOffThread → ManagedThreadBase::KickOff → ... into JIT-ed frames) —
  i.e. a managed thread continued running in the fork child, which is undefined post-fork state.
• The parent Worker waits forever; the job shows the current step "in progress" indefinitely.
• On job cancellation / 60-min timeout the spinning fork child survives, so the runner stays
  wedged (pegged core, listener busy) until the processes are killed manually. Reboot doesn't
  help — the next job hits the same spawn deadlock.

To Reproduce

1. Self-hosted runner v2.336.0 on macOS 26.5 (arm64), installed as a launchd service (./svc.sh install).
2. Trigger any workflow job with a few consecutive steps (checkout + bash steps).
3. Within seconds, a step never starts its process; job hangs until timeout.

Hit rate on our machine: 6/6 jobs after updating; 0 occurrences on 2.335.1 before and after (rolled back via the on-disk bin.2.335.1/externals.2.335.1 + disableUpdate: true).

Expected behavior

Process spawn completes (or fails); no fork-child spin; cancellation kills the worker cleanly.

Runner Version and Platform

• Runner: 2.336.0 (self-update from 2.335.1 on 2026-07-22)
• OS: macOS Tahoe 26.5.2 (25F84), arm64 (Apple M5 Pro)
• Service: launchd user agent via svc.sh

Job Log Output

[...] INFO ProcessInvokerWrapper] Finished process 21919 with exit code 0, and elapsed time 00:00:00.3804644.
[...] INFO ProcessInvokerWrapper] Starting process:
[...] INFO ProcessInvokerWrapper]   File name: '/bin/bash'
[...] INFO ProcessInvokerWrapper]   Arguments: '-e /Users/.../_work/_temp/<uuid>.sh'
[...] INFO ProcessInvokerWrapper]   Force kill process on cancellation: 'False'
<nothing further; only HostContext heartbeat lines every 10s>

sample output of the spinning fork child available on request (full stack).