core: MPS manager can kill the same user's other GPU jobs #166

Problem

_mps_daemon_running uses a host-wide pgrep, so a daemon serving another job's pipe directory looks like ours; the probe of our own pipe dir then fails and the "stale" cleanup runs pkill -u $UID on every MPS server and control process.

Evidence and reproduction

opt_prime/opt_prime/mps_manager.py:504-530, :801-808. Case: job A runs with /tmp/mps-a, job B starts with the default dir, job A's daemon is killed. Verified by reading.

Environment

optimus-rm @253be20; Python 3.10, torch 2.10.0+cu128, transformers 5.1.0; RNGD+A100 host (CPU / one A100 for the checks).

Scope

mps_manager.py.

Done when

Daemon liveness is decided by our own control socket only; no host-wide pkill; a test with a fake pipe dir.

Owner

@gspark-etri