[CodingInAVan]

Summary

The agent decides that a producer finished by probing its .gpufl-session.lock. The probe opened the lock read-write. An agent running as another user (for example the systemd user gpufl-agent) usually has no write access to it, so the open failed, the state was UNKNOWN, and the session never completed. An H100 run needed chmod g+rw on every lock after each run.

• Open the lock read-only and try a shared lock. It still conflicts with the writer's exclusive lock (fcntl on POSIX, LockFileEx on Windows), so ACTIVE and UNOWNED mean what they did.
• When the lock cannot be opened, log one warning per session directory naming the permissions the agent's user needs.
• Add a README with a section on running the agent as a separate user: the setgid session directory, the producer's umask, and the chmod still needed for clients up to 1.3.0rc6.

Pairs with gpu-flight/gpufl-client#177, which creates the lock group-readable (0660 masked by the umask). Each change is safe without the other: with an older client the agent still cannot open the lock and now says so.

Testing

• ./gradlew check passes: 176 tests, 1 skipped (a POSIX-only test on this Windows host), coverage gate passed.
• New: a child JVM holding an exclusive lock on byte 0 (the Windows client's range) probes ACTIVE, then UNOWNED after it exits; and, on Linux/macOS, a mode-400 lock probes UNOWNED while mode 000 probes UNKNOWN. The POSIX test runs first in CI.