Bug: CUDA MPS CLI tests fail on macOS #2510

Summary

tests/training/test_cuda_mps_cli.py is not reliably runnable on macOS. With the normal pytest temporary directory, make test-all reaches 2673 passed / 133 skipped and then fails 13 tests in this file. Shortening the temporary root with TMPDIR=/tmp removes the AF_UNIX path failures but still leaves 4 failures.

Area

CI/tooling

Reproduction

Normal macOS pytest temporary root:

make test-all

Focused run with the longer default Darwin pytest temporary root:

uv run pytest -m 'not slow' tests/training/test_cuda_mps_cli.py -q

Focused run with a shorter temporary root:

TMPDIR=/tmp uv run pytest -m 'not slow' tests/training/test_cuda_mps_cli.py -q

Expected behavior

The MPS CLI test module either passes on supported test platforms or explicitly skips macOS so the repository test gate can complete locally. Any test that intentionally fakes a Linux/NVIDIA host should exercise its intended validation branch independently of the physical host platform.

Actual behavior

Default Darwin pytest paths fail before the intended assertions because Unix-domain socket paths exceed the platform limit. With TMPDIR=/tmp, four tests still fail because platform faking is missing from the direct start_daemon() test paths, so production _require_linux_nvidia() fails first.

Default-path failures (13):

test_start_rejects_task_per_gpu_until_lease_contract_lands
test_status_reports_managed_control_pipe_and_diagnostic
test_start_daemon_uses_explicit_directories_and_records_identity
test_start_daemon_waits_until_control_answers_query
test_start_daemon_cleans_up_unrecorded_control_after_readiness_failure
test_start_daemon_recovers_only_inert_failed_control_artifacts
test_start_daemon_failure_reports_control_log_and_cleans_artifacts
test_default_runtime_recovers_orphan_control_sockets
test_start_daemon_refuses_existing_unmanaged_control_path
test_start_daemon_quarantines_stale_record
test_doctor_uses_sole_live_managed_daemon_pipe
test_doctor_reports_missing_control_daemon_and_exits_invalid
test_regular_file_control_pipe_is_rejected_by_probe

Short-path failures (4):

test_start_rejects_task_per_gpu_until_lease_contract_lands
test_start_daemon_uses_explicit_directories_and_records_identity
test_start_daemon_waits_until_control_answers_query
test_start_daemon_cleans_up_unrecorded_control_after_readiness_failure

Platform guard mismatch example:

CudaMpsError: CUDA MPS management requires Linux and NVIDIA CUDA; current platform is 'Darwin'.
Expected regex: 'current support is single_gpu'

AF_UNIX/path-length examples:

OSError: AF_UNIX path too long
CudaMpsError: MPS control socket path is too long (99 bytes; maximum 95): /private/tmp/pytest-of-yves/.../control

Environment

• OS: macOS (Darwin)
• Python: 3.13.12
• Commit: ed9ff2f9 base (also observed with b1a623c0, which does not touch this module)
• Backend: MuJoCo environment installed; tests invoke the CUDA MPS CLI with fake subprocesses

Logs and artifacts

• Full local run discussed in PR fix(rl): initialize off-policy MPS inference timer #2509: 13 failed, 2673 passed, 133 skipped.
• Focused short-root run: 4 failed, 39 passed.
• Earlier AF_UNIX fix: closed issue macOS MPS tests exceed AF_UNIX path limit #2089 and commit a1084ce8 (default runtime path), but the current test fixture and several direct tests still derive paths from pytest tmp_path.
• Direct start_daemon() tests lack the linux_host fixture's platform.system() fake where they call helper logic directly instead of through the fixture path.

Proposed fix direction

• Keep AF_UNIX-producing tests on a short shared test-only temporary root, or make the module skip when the effective socket path cannot fit the platform limit.
• Apply the fake Linux platform consistently to direct start_daemon() tests that intentionally simulate a Linux/NVIDIA deployment.
• Keep production _require_linux_nvidia() behavior unchanged; this is a test-environment issue only.