The previous test asserted that no two adjacent samples from a
~10 000-element space were equal across 100 iterations. The birthday
math gives that adjacency check a ~1 % flake rate per run; with the
new gitea.lan CI volume that turned into observable random failures
(go-unit #51 on feature/enable-actions-cache hit "Should not be:
'6635'").
Replace adjacency with a distinctness floor over a wider 200-sample
draw. A stuck generator (single value) lands at 1 unique; a
256-element range lands at ~196; the natural full-range generator
lands at ~198. A floor of 150 catches the failure modes the test was
actually written to guard against and never trips on legitimate
randomness.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Ilia Denisov