feat(cli): manage user-owned CUDA MPS daemons

[TATP-233]

Driving issue

Closes #2079.

Final UX

The MPS lifecycle is a short, flat command. On the supported single-GPU host, no arguments are required:

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

Read-only host inspection:

uv run uni-cumps status

Explicit overrides remain for multi-GPU selection, stable labels, and service-owned paths:

uv run uni-cumps doctor --gpus <gpu-index-or-uuid>
uv run uni-cumps start \
  --gpus <gpu-index-or-uuid> \
  --name <name> \
  --pipe-dir /absolute/path/mps/pipe \
  --log-dir /absolute/path/mps/log
uv run uni-cumps env --name <name>
uv run uni-cumps stop --name <name>
uv run uni-cumps stop --all

Defaults

Setting	Default
GPU	the sole visible physical GPU, resolved to its canonical UUID
daemon name	gpu-<uuid-prefix>
pipe directory	~/.cache/unilab/cuda-mps/<name>/pipe
log directory	~/.cache/unilab/cuda-mps/<name>/log
env / stop target	the sole live UniLab-recorded daemon
execution mode	detached control daemon

Ambiguity does not silently select a target:

• more than one visible GPU requires --gpus;
• more than one live daemon requires --name;
• no live daemon is an error, not an unmanaged-pipe fallback.

Delivered behavior

• Adds the uni-cumps package entrypoint, owned by src/unilab/training/cuda_mps_cli.py.
• Implements status, doctor, start, stop, stop --all, and env.
• Resolves selectors to canonical physical GPU UUIDs and rejects MIG UUIDs.
• Models single_gpu, single_task_multi_gpu, and task_per_gpu.
• Implements only single_gpu; multi-GPU/all daemon execution fails closed and points to CUDA MPS under multi-GPU DP: evidence gaps and design constraints #2063/roadmap: multi-GPU tensor runtime scaling (M10) #1672.
• Stores records under ~/.cache/unilab/cuda-mps/<name>/daemon.json.
• Scopes records by (UID, hostname, name).
• Persists GPU UUIDs, absolute pipe/log paths, control PID, and Linux /proc/<pid>/stat start-time.
• Parses procfs PID/start-time after the parenthesized process name, so names such as (nvidia-cuda-mps) do not break stop authority.
• Preserves the exact nvidia-smi UUID spelling, including hyphens and case, for CUDA visibility masks.
• Permits stop only when UID, host, record, PID, and start-time all match.
• Quarantines stale records; never attaches to or stops unmanaged pipes.
• doctor selects the sole matching live managed daemon and probes its pipe/log environment.
• Prints launcher environment without mutating the parent process.
• Adds flat subcommand/flag shell completion.
• Adds ADR-0014 and narrows the ADR-0013 supersession to daemon lifecycle ownership.
• Updates English and Chinese CLI/production documentation.
• Keeps trainer semantics unchanged: training.cuda_process_sharing=mps remains explicit, with no auto mode and no implicit daemon selection.

Impact

• Training runtime semantics: unchanged.
• Backend scope: doctor reuses the existing MJWarp/single-rank fail-closed probe contract.
• Platform scope: daemon lifecycle targets Linux/NVIDIA CUDA; status remains useful and fail-closed elsewhere.
• ROCm profile retains the new entrypoint through a scripts parity test.

Validation

Commands run on final local head 52220875:

make check
make test-all

Results:

• formatting and lint passed;
• mypy: 133 source files, no issues;
• pyright: 0 errors, 0 warnings;
• non-slow coverage run: 1847 passed, 42 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/test_completion.py \
  tests/scripts/test_rocm_dependency_profile.py tests/scripts/test_check_docs.py -q

Result: 69 passed.

Sphinx strict build was run successfully with no warning introduced by this change.

Real Linux/NVIDIA lifecycle smoke

Completed on the available host with a real nvidia-cuda-mps-control:

uv run uni-cumps start --name finale2e \
  --pipe-dir /tmp/final-e2e/pipe \
  --log-dir /tmp/final-e2e/log
uv run uni-cumps doctor
uv run uni-cumps status
uv run uni-cumps env --name finale2e
uv run uni-cumps stop --name finale2e
uv run uni-cumps status

Observed:

• start wrote a record containing PID 2578762, procfs start-time 131179025, canonical UUID, UID, host, and absolute pipe/log paths.
• doctor selected the live managed daemon and reported valid=True.
• status reported the daemon live.
• env emitted the correct CUDA_VISIBLE_DEVICES, pipe, and log exports.
• stop returned 0 and removed the record.
• final status reported managed_daemons=none; the control process was no longer present.

Regression coverage added from this real run:

• parenthesized procfs comm parsing;
• PID mismatch rejection;
• canonical UUID spelling with hyphens/case preserved;
• doctor selection of the sole live managed daemon and its pipe/log environment.

Scope boundaries / limitations

• single_task_multi_gpu and task_per_gpu are contract names only; execution remains blocked pending CUDA MPS under multi-GPU DP: evidence gaps and design constraints #2063/roadmap: multi-GPU tensor runtime scaling (M10) #1672 and a lease/reference model.
• Doctor reports the first blocking prerequisite through the existing probe contract, not every possible subsequent error.
• Trainer does not automatically select or lease a daemon.
• No multi-node, MIG, SM percentage, memory-limit, or affinity support is introduced.
• Long-duration daemon soak is not part of this PR, but the real start/status/env/doctor/stop lifecycle has now been exercised once on Linux/NVIDIA.