Set up job called twice and fail immediately (race-condition?)

[Kaspik]

Describe the bug
When running actions in matrix where resources are far lower than requested count, jobs fail immediately due to set up job step conflicting on the same runner.

We use self-hosted macOS Bare metal machines with one runner per machine.

To Reproduce

• set matrix to run 30 jobs
• set max-parallel to 4
• push to PR and see

Expected behavior
4 actions should run with another 4 following once some complete.

Note: Re-running usually ends up the same on different self hosted runner, and we get stuck on the issue for random amount of time till Github action decide to ... go through,

Runner Version and Platform

2.330.0, macOS 26.2

What's not working?

Runner and Worker's Diagnostic Logs

_diag is empty after the run finishes.