ci: run the Python contract suites

[aurascoper]

Stacked on #36 — merge that first. test_session_consume imports eeg_spectral, which cannot import on numpy>=2 without it, so this job would be red on its own.

Prerequisite for #33 going ready, per the agreed gate github_ci_runs_tests_eval: true.

Why nothing caught the missing module

build-and-test is Swift-only. Neither NeuralComposeEEG/tests nor Tests/eval has ever run in CI, leaving 62 Python tests ungated — which is exactly how #33's test could import an uncommitted module and still show a green check.

pytest is required here, not preferred

Tests/eval mixes four harness styles. Most modules are bare def test_* functions with no TestCase and no __main__ runner:

module	python -m unittest collects	real tests
test_dialectic_corpus_quarantine	2	2
test_eeg_spectral_compat	6	6
test_eeg_channel_quality	0	6
test_embedding_space	0	9
test_eval_stats	0	8
test_overnight_watchdog	0	8
test_state_reconstruction	0	4
test_state_trajectory_analysis	0	5

unittest prints OK for every one of those zero-collection modules. My first draft of this job used unittest and would have shipped a green check that ran 8 of 62 tests. pytest collects all 62.

Exclusions, each with a reason

test_joint_embedding.py imports mlx.core (Apple Silicon only) — run locally.

Three cases are deselected because they already fail on this base, on both NumPy 1.26.4 and 2.4.6. They are not caused by this PR and are not silently dropped — see the tracking issue. The remaining tests in those files still run.

Result

NeuralComposeEEG contracts   33 passed   (unittest discover; grows to 46 once #35 lands)
Tests/eval                   59 passed, 3 deselected

torch installs from the CPU wheel index — the default Linux wheel is a ~800MB CUDA build for a runner with no GPU.

[aurascoper]

Tracking issue for the three deselected cases: #38. It records which are wrong tests and which is a real defect — briefly: test_mann_whitney_u_disjoint asserts a p-value that is unattainable at n=3 per group, test_cka_independent ignores biased-CKA inflation when d ≫ n, and cohens_d genuinely returns 0.0 (i.e. "no effect") for two perfectly separated zero-variance groups.