fix(cli): keep default MPS runtime within socket limits

[TATP-233]

Driving issue

Closes #2091.

Problem

The zero-argument uni-cumps workflow could strand itself after a failed or exited NVIDIA control daemon:

• the old default pipe path under a normal home directory could exceed NVIDIA's effective Unix-socket boundary;
• NVIDIA returns exit code 1 with no stderr when the control path is too long;
• failed launches leave control_lock and a pipe/log FIFO;
• exited control daemons leave control and control_privileged sockets;
• the next start then reported “Refusing to attach to an existing control path,” while env/stop had no daemon record to use.

Measured on CUDA MPS 13.0 / driver 580:

• 95-byte control paths succeed;
• 96-byte and longer paths fail with no stderr.

Fix

Default runtime paths are now short and private:

/tmp/uni-cumps/<daemon-name>/pipe
/tmp/uni-cumps/<daemon-name>/log

All default runtime directories are created with mode 0700 and owned by the invoking user.

The default daemon name now uses the complete canonical GPU UUID:

gpu-<full-uuid>

rather than a shortened prefix, avoiding daemon-name collisions.

Explicit --pipe-dir values are checked before launch. A resulting control path over 95 bytes fails with an actionable message recommending a short path such as /tmp/<short-name>/pipe.

Failed-launch recovery is narrow:

• known inert artifacts (control_lock and the pipe/log FIFO) are removed after a failed control launch;
• in the UniLab-owned default runtime only, orphan control / control_privileged sockets are removed when no PID file exists;
• explicit user-supplied directories are never modified;
• live sockets with daemon PID evidence are never removed.

A nonzero control exit now appends the control-log tail, making silent NVIDIA failures diagnosable.

Final workflow

Verified on the available Linux/NVIDIA host:

uv run uni-cumps start
eval "$(uv run uni-cumps env)"
uv run uni-cumps doctor
uv run uni-cumps stop

All four completed successfully; doctor exited 0. A second start while live correctly reported the existing live daemon. Restarting after a dead control process also succeeded after safe orphan cleanup.

The immediate start → doctor pair was additionally repeated 20 times in one shell; all 20 doctors returned valid=True and exit 0. The orphan check now validates PID liveness and the /proc/<pid>/stat process name rather than treating the mere presence of a stale PID file as live ownership.

Validation

Commands run on final local head a1084ce8:

make check
make test-all

Results:

• ruff/mypy/pyright passed;
• non-slow coverage run: 1845 passed, 43 skipped, 488 deselected;
• benchmark smoke: module-mode 34/34, script-mode 35/35.

Focused checks:

uv run pytest tests/training/test_cuda_mps_cli.py tests/scripts/test_check_docs.py -q

Result: 57 passed.

New coverage includes:

• full-UUID default names;
• explicit socket-path length rejection;
• removal of inert failed-launch artifacts while preserving unrelated files and live sockets;
• orphan default-runtime socket cleanup, including stale PID files and procfs process-name matching;
• control-log diagnostics on nonzero exit.

Scope boundaries

• No trainer lease/selection behavior is introduced.
• No multi-GPU, task-per-GPU, MIG, SM-percentage, memory-limit, affinity, or multi-node support is claimed.
• Explicit pipe directories are never cleaned by UniLab.