The test was flaky because it used crypto/rand for random data,
causing non-deterministic chunk boundaries. With small sample sizes
(100KB / 8KB avg = ~12 chunks), variance was high - sometimes only
42.9% overlap instead of expected >50%.
Fixed by using math/rand with seed 42 for reproducible test results.
Now consistently achieves 91.7% overlap (11/12 chunks).
Alexander Renz
b2a78cad2a