test(eval): use an attainable exact Mann-Whitney significance case

[aurascoper]

Second of three PRs resolving #38. Stacked on #39 — merge that first, then this retargets to docs/eeg-methods-scope.

No production change

The test asserted a p-value its own sample size makes unreachable:

mann_whitney_u([1, 2, 3], [10, 11, 12])   # p = 0.1, asserted < 0.05

With n₁ = n₂ = 3 there are only C(6,3) = 20 arrangements, so the smallest attainable two-sided p is 2/20 = 0.1. Completely disjoint groups return exactly that. No correct implementation can pass this assertion at that sample size. The implementation was right; the test was impossible.

Four per group makes the claim attainable

n=3 per group   U=0.0  p=0.1000     (exact minimum 2/20)
n=4 per group   U=0.0  p=0.0286     (exact minimum 2/70)

Both verified against the real implementation. The assertions now also pin u_statistic and both sample sizes, so a future change to the return shape can't slip through on the p-value alone.

Result

test_eval_stats 10 passed locally. Removes the second of three deselections.