fix(gateway): harden macOS PID detection

[BrennerSpear]

Summary

This hardens macOS gateway PID discovery so Hermes stops producing false negatives when a launchd-managed gateway is alive.

Specifically, this PR:

• parses both legacy tabular and newer plist-style launchctl list <label> output in _get_service_pids()
• uses macOS-safe ps arguments for manual gateway process discovery while preserving the existing non-macOS path
• adds regression coverage for plist-style launchctl output and the macOS process-scan path
• updates an existing PID-file fallback test to match the macOS-safe ps invocation

Problem

On the affected macOS install, these commands disagreed:

• hermes gateway status
• hermes cron status

launchctl showed the gateway loaded and running with a real PID, but find_gateway_pids() returned no processes, so hermes cron status incorrectly claimed the gateway was not running.

Root cause

There were two macOS-specific detection bugs:

1. _get_service_pids() only understood older tabular launchctl list <label> output, but newer macOS returns plist-style output like:

   {
       "Label" = "ai.hermes.gateway";
       "PID" = 7817;
   };
   

2. _scan_gateway_pids() used ps -A eww -o pid=,command= for every non-Windows platform. That works on Linux, but the macOS ps on the repro machine rejects that invocation.

What changed

launchd PID parsing

Added _parse_launchctl_pid() and used it from _get_service_pids() so Hermes now accepts both:

• legacy tabular launchctl list output
• plist-style launchctl list output from newer macOS

macOS-safe process scanning

_scan_gateway_pids() now chooses:

• macOS: ps -A -ww -o pid=,command=
• non-macOS: existing ps -A eww -o pid=,command= path

Tests

Run locally:

• pytest tests/hermes_cli/test_gateway.py tests/hermes_cli/test_update_gateway_restart.py -q

Result:

• 61 passed

Notes

I found two existing open PRs in similar territory (#11293 and #10636). I'm opening this because I reproduced the bug locally against a live macOS launchd install, implemented the fix against current main, and validated it in the current test layout.

[BrennerSpear]

Closing this as a duplicate of #11293.

I compared the overlapping PRs after opening this and #11293 is the best existing upstream thread for this bug: same root cause, narrower scope, and good regression coverage. I left a supporting comment there with the local macOS repro details from my install.

Keeping discussion and maintainer attention on that PR.