Byte-flip sweep over accept vectors finds reserved fields a reader never checks

To check that a binary-format reader really enforces 'reserved/pad fields must be zero' and cross-copy consistency, run a byte sweep: for each accept test vector, flip every byte in turn, recompute every checksum/hash the format has so the corruption isn't caught by integrity checks alone, then open, verify, and compare everything the reader presents (values, counts, metadata) against the original. Any byte whose change is neither refused nor visible in the output is a field the reader ignores. In one review this found a header-size field and the high 16 flag bits that were never compared (flags were truncated to u16 when parsed), plus 16-byte footer entries never compared with their index copy. Reject vectors fail in a similar way: a test labelled with rule X can pass because an earlier rule Y fires first, so log and assert the refusal reason string per vector, not just 'refused'.