fix(eval): stop zero-variance Cohen's d from reporting no effect

[aurascoper]

First of three PRs resolving #38. Removes exactly one CI deselection.

The defect

cohens_d returned 0.0 whenever pooled SD was zero, regardless of whether the means were identical or completely separated:

cohens_d([1, 1, 1, 1], [5, 5, 5, 5])  # 0.0 — "no effect" for total separation

Zero pooled variance now distinguishes the two cases:

input	before	after
equal constant groups	0.0	0.0 (a true zero)
separated constant groups	0.0	NaN (undefined)
under-sized samples	NaN	NaN (unchanged)

NaN over infinity, per the issue: d is undefined without within-group variation, NaN doesn't pretend the effect is absent, it matches the function's existing under-sized-sample behaviour, and it forces callers to treat the estimate as degenerate rather than ranking it as an ordinary finite effect.

Consolidation, because there were four copies

eval_stats.py's docstring already claims it consolidates cohens_d "previously duplicated in embedding_analyze.py and statistical_analysis.py" — but the duplicates were never removed, and analyze_results.py held a fourth. All four were byte-identical in logic, so fixing one would have left three wrong ones.

analyze_results.py, embedding_analyze.py, and statistical_analysis.py now import the shared implementation, following the sys.path.insert(0, str(Path(__file__).parent)) idiom already used in this directory for bootstrap_ci. Verified at runtime that all three resolve to the same function object.

Scope note: mann_whitney_u is also duplicated across these scripts. Left alone deliberately — that is 38-B's subject and does not belong in this diff.

Tests

test_cohens_d_large_effect conflated "large" with "zero-variance degenerate". Replaced by three tests that separate the cases:

test_cohens_d_identical_constant_groups_are_zero
test_cohens_d_separated_constant_groups_are_undefined
test_cohens_d_large_nondegenerate_effect

Expected CI counts

The issue projected 84 passed, 2 deselected. Actual will be 88 collected / 2 deselected / 86 passed: this PR replaces one test with three, so the collection grows by two. Nothing was skipped — the progression is 83 → 86 passed because the suite gained tests alongside the fix.

Local: test_eval_stats 9 passed, 1 failed (the still-deselected Mann-Whitney case, 38-B's subject). py_compile, YAML parse, and git diff --check all clean.