The integer-synthesis decoder (wai.neural.int_synth) is the one neural path
that can make WAI's promise — bit-identical output on every machine, no
tolerance — because it is pure i64 arithmetic, not float ONNX. It had a
constructed-in-test golden but no committed fixtures and was not in the
pure-Rust conformance matrix (it was buried in the ort-gated `neural`
module, so it couldn't build without ONNX Runtime).
Extract it to the crate root behind a light `neural_int` feature (serde_json
+ blake3, no ort): self-contained IntSynthError/IntImage, identical i64
decode. `neural` now pulls neural_int in and keeps a thin shim mapping onto
its DecodedImage/DecodeError, so the ort-backed dispatch is unchanged. The
byte-exact-conformance crate's #[path] include follows the move (its
independent constructed golden still holds — a second implementation path).
Add wai_int_synth_vectors (gen/verify) + a committed neural-conformance/
corpus: three realistic multi-layer decoders (distinct shapes, integer-ReLU
stacks), authored from a fixed SplitMix64 sequence, each frozen to a
domain-prefixed BLAKE3 of its decoded RGB. The wai-conformance workflow runs
the verify on the four-platform matrix (x86-64 + arm64 across Linux, macOS,
Windows) with no ONNX Runtime present — the integer-exact contract proven on
real silicon, every push. The float decoders are deliberately excluded:
their ONNX synthesis is float, so a byte-equality corpus would be a false
promise for them.
Co-Authored-By: Claude Opus 4.8 <[email protected]>
293911e992