fix(test): [T176] decisive normality display fixture — stop cross-version decision flip (#275)

Follow-up to the goldens robustness fix: the LTS (1.10) CI cleared 8 of the 9
mismatches but left `normality`, whose golden pins the H₀ DECISION column
("Reject" / "Fail to reject"). That word is numerically derived (p-value vs 0.05)
and cannot be masked like a number/star — and the underlying `randn` stream is not
stable across Julia versions, so the Gaussian fixture's near-boundary p-values
flipped their decisions 1.10↔1.12.

Rather than mask decision words globally (which would needlessly weaken the
johansen/adf goldens, whose decisions are decisive), make the ONE borderline
fixture decisive: feed `randexp` (skewness ≈ 2, excess kurtosis ≈ 6) so every test
in the battery rejects with a statistic in the thousands (p ≪ 0.001). No BLAS/OS/
version last-ulp difference can move a statistic of ~7300 below its ~7.8 critical
value, so the decision column is now identical everywhere. Same principle as the
high-SNR coefficient fixtures whose significance is unambiguous.

Verified: goldens 20/20 + invariants 124/124 green on 1.12; all 7 decisions now
"Reject". Only normality.txt regenerated.

Claude-Session: https://claude.ai/code/session_01QZizX88NDsBZrBgdqPcLQp
