Exercise the CUDA comparability warning

[emrefbulut]

What changed

Two tests covering the CUDA comparability warning in iqforge train, which no test had ever executed.

Why

README.md and CONTRIBUTING.md both promise it:

A CUDA run prints a warning that its numbers are not bit-comparable with CPU runs

The line exists (src/iqforge/cli.py:1055) and nothing ran it. It is the only thing that tells a CUDA user their numbers cannot be placed next to a CPU table, and a promise nothing exercises is a promise nobody has checked.

How it was verified

No GPU needed, and none wanted. torch.cuda.is_available is pinned True so resolve_device("cuda") and describe_environment run for real and produce a genuine cuda stamp; train_baseline is replaced so nothing reaches for a CUDA context this machine does not have. Mocking availability is not the same as having a device — manual_seed_all and Adam would find that out, which is why the existing test_models.py tests stop at placement too.

What is under test is the CLI branch reading result.environment["device"] — the line that decides whether a CUDA user is told anything.

Both halves of the branch are covered. A warning printed unconditionally would satisfy the positive test while telling every CPU user their numbers are not comparable, so the CPU case asserts its absence.

Mutations, each red on the right test:

mutation	caught by
delete the warning (if False:)	test_a_cuda_run_warns_that_its_numbers_are_not_comparable
print it unconditionally (if True:)	test_a_cpu_run_prints_no_comparability_warning
ignore --device, always pass cpu	test_a_cuda_run_warns_that_its_numbers_are_not_comparable

uv run pytest tests/test_cli_output.py -k comparab -q   # 2 passed
uv run ruff check .                                     # clean

One measurement worth recording. The assertions compare with all whitespace removed. rich wraps at a console width this test cannot set, and it strips the space it wrapped on — which turns "Do not mix" into "Do notmix". The existing _flat helper drops newlines but does not restore that space. Found by the assertion failing, not predicted.

Noticed, not fixed here

Two gaps in the same feature, both out of scope for a test-only PR:

1. measure-leakage --device cuda prints no such warning. Only train has it. A CUDA leakage measurement emits environment: device cuda | ... and nothing else — and that command is the one whose output gets pasted into a table.
2. The warning sits after the empty-test-split early return (if result.test_accuracy is None: return), so a CUDA run with an empty test split trains on the GPU and exits without warning.

Conventions it touches

• 2 A passing test does not prove it can fail — three mutations, each confirmed red on the test that should catch it.
• 3 Do not claim what you did not measure — the whitespace behaviour above was measured, and the two gaps are reported rather than left for someone to discover.