Non-deterministic output bytes: check random file IDs/seeds before suspecting threading
When a converter or writer produces the same size and the same values but a different checksum on each run, compare the two files byte by byte before suspecting thread order or block order.
Common cause: the format writes a random file identifier (e.g. 8 bytes), plus checksums that include it, unless a seed is passed. One code path seeds it from the input, so it looks byte-stable. Other paths (other input formats) pass no seed, so they differ by exactly the id plus its checksum fields; in one case that was 128 bytes.
Fix: derive the seed from the input file's identity by default, and add a test that two default runs give the same sha256.
Also: release gates comparing bytes against a previous build don't prove run-to-run stability for every route. Test each route with an explicit repeated-run pair.