D070 - The hasher read the digest backwards, from the very first commit

decided · 2026-08-19

NidHasher took the first eight digest bytes little-endian. It should be big-endian. Every NID this project has ever computed was wrong.

Why nothing caught it. Every test hashed a name with an arbitrary suffix and compared the result against its own output. That is self-consistent, and self-consistent with a wrong byte order. The decoder was right and the hasher was wrong, and the two never met - decode_symbol_name's own documentation admitted as much, saying the choice was "consistent but not independently verified".

What exposed it. Hashing published C library names against a real import table. The right convention matched 66 of 468; every other combination of byte order and suffix placement matched zero. There is no ambiguity in a signal like that.

The invariant that was missing, now a test: a hash must survive a round trip through the symbol-name encoding. decode(encode(hash(name))) == hash(name). That needs no suffix, no real module, and no external reference - it compares the two halves of this crate against each other, which is exactly what nothing was doing. encode_nid exists mainly so that test can.

The general lesson, worth more than the fix: a test that compares a component to itself proves only internal consistency. Both halves of a two-sided convention can be individually well-tested and jointly wrong. Where two representations must agree, test the agreement.