We analyzed the data on Slack, majority of differences can be attributed to the fact that zstd compression is in use and it does not always guarantee byte-by-byte identical outputs for the same input. Assuming hostA and hostB had different CPU characteristics (number of cores, etc.) that explains most of the difference.
There’s however the difference in splitter output too, which we’re still trying to get to the bottom of: all split points except first 3 are the same, but it appears that first 6291640 bytes got split differently (1685450+2132853+2473337) vs (2097152+2097152+2097336).