2026-08-19 (evening) - Names, and proving they are ours

The name search went from built to working, and then grew the provenance machinery it needed. D070-D073; 24 crates, 265 tests.

Two bugs and a milestone.

The hasher had been reading the SHA-1 digest little-endian since the first commit. Every NID this project ever computed was wrong. Nothing caught it because every test hashed with an arbitrary suffix and compared against its own output - self-consistent, and self-consistently wrong.

The suffix I had been treating as unobtainable was fine all along; the byte order was the problem. Once fixed:

Executable Imports Named
2 MB 715 83
96 MB 1,380 161

And the function the 96 MB executable calls 431 million times is sceKernelDirectMemoryQuery. It asks about direct memory, is told "unimplemented", and asks again forever. The wall has a name.

Surprises

Outstanding

88% of imports are still unnamed. That is a vocabulary problem, not a method problem - symbols/wanted.txt is the work list and extending data/vendor.toml needs no rebuild.