A.merge(B) reports 100 for every percentile while B.merge(A) reports 10,000 for every percentile; the exact pooled values are p50=100, p95=10,000, p99=10,000 over 1,060,000 observations, and both orders report a wrong count.
Python

Merging reservoir samples by replaying one buffer into the other makes percentiles order-dependent

/

A.merge(B) reported p99=100, B.merge(A) reported p50=10,000 — both wrong. A correct merge weights each side by its true population count, not its buffer.