ci(d0): run Tests/eval non-vacuously, and fix what that surfaced

[aurascoper]

head:    b8c57af56b8cfa426097bf555f323101c64f9eb0
base:    main@859feaa        # the D0-R merge commit
lane:    D0  (CI substrate; precedes A0/A → B)
files:   8
commits: 2

The finding

Tests/eval has seven modules. Exactly one ran in CI — test_adr_references.py, via D0-R's direct step. The other six executed nowhere, on any branch, ever.

Turning them on required fixing four things they were hiding.

1 · The harness could not have worked

Six of seven modules are bare def test_* with no unittest.TestCase, so python -m unittest collects zero from them — and unittest discover fails outright, since Tests/eval has no __init__.py. Either path reports success while running nothing. pytest is forced here, not stylistic.

2 · An import-time crash

_trapz = getattr(np, "trapezoid", getattr(np, "trapz"))     # Scripts/eeg_spectral.py

Python evaluates that default before the lookup it is a fallback for. On numpy 2.x — which removed trapz — it raised AttributeError at import, on precisely the version the fallback existed to support. Three further modules failed collection on a missing scipy, a real transitive dependency now installed rather than assumed.

3 · Three assertions that had never run

test	verdict
cohens_d([1,1,1,1],[5,5,5,5]) → 0.0	code defect. Zero pooled SD means both samples are constant; this reported no effect for the most extreme separation representable. The limit is signed infinity; identical constants are the only genuinely zero case. Fixed in all four copies.
mann_whitney_u([1,2,3],[10,11,12]), p < 0.05	impossible test. At n=3 vs 3 the exact two-sided floor is 2/C(6,3) = 0.1. No implementation could have passed. Now asserts the floor — which also pins that the implementation is exact rather than normal-approximating — plus a companion at n=4 where 2/C(8,4) ≈ 0.0286 clears 0.05.
cka(randn(50,128), randn(50,256)), < 0.3	wrong estimator. Biased linear CKA inflates when d ≫ n; independent data scores ~0.78. < 0.3 describes an unbiased estimator this code does not implement. Now bounded with the reason recorded, plus a companion showing the inflation shrinks as n grows.

Consolidating the four cohens_d copies is deliberately not done here — that structural change belongs to the lane that owns it.

4 · Guards, because a job that runs nothing still exits 0

collection errors == 0     a module that stops collecting fails loudly
collected >= 55            floor below the current 63

Proven non-vacuous: reintroducing the trapz bug yields 1 collection error and drops collection to 51 — both gates fire.

NeuralComposeEEG/ is absent on main until lane B. The guard says so explicitly and exits 0; once present, zero collected tests is an error.

Evidence

Tests/eval          30 collected + 4 errors  →  65 passed, 0 failures
trapz mutation      1 collection error, 53 collected — both guards fire
ADR gate (D0-R)     16 tests OK, PASS
swift build         unaffected
git diff --check    clean
CI                  build-and-test + python-eval SUCCESS

5 · Statistical-contract fixes (second commit)

Review of the first commit found two defects in functions this job now executes
for the first time. Both shared a shape: the code was more general than its
tests, and the untested region was exactly where it was wrong.

• Cohen's d pooled variance is df-weighted for unequal sample sizes.
  sqrt((var_a + var_b) / 2) matches the conventional estimator only at equal
  n, and the function accepts arbitrary lengths — every earlier test used
  equal groups. On n=3 against n=9 the reported effect moves 2.540 → 2.033.
  The regression test asserts toward the weighted value and away from the
  unweighted one, so reverting fails rather than silently changing a number.

• ±Infinity for unequal constant groups is an explicit project policy, not
  the definition of Cohen's d. Persistence callers must map non-finite values at
  JSON/CSV boundaries: JSON has no Infinity literal (json.dumps emits a bare
  Infinity that strict parsers reject) and CSV has no convention.

• Mann-Whitney uses exact for tie-free samples within a stated size cap,
  asymptotic otherwise. Previously it omitted method= and inherited SciPy's
  auto, whose rule has changed across releases and would move a reported
  p-value with no commit here. The earlier claim that the test pinned an exact
  p-value was true of the value but not the mechanism; it is true of both now.

  n=3 exact floor   0.100000   = 2/C(6,3)
  n=4 exact         0.028571   = 2/C(8,4)
  with ties         0.103754   asymptotic, as required
  

• CI bounds numpy>=2,<3 and scipy>=1.14,<2. The trapz fix is
  numpy-2-specific and the Mann-Whitney contract is SciPy behaviour, so an
  unconstrained install could change a p-value with no commit here.

test_joint_embedding.py is the one exclusion, by name and reason: it imports mlx (Apple Silicon only) and needs a converted model.

Independent of #44, which stays parked for review.