typesafe: native ONNX runtime path is TODO, no macOS prebuilt, and README onnx entry points fail on WASM

[stuinfla]

Thanks for shipping @ruvector/typesafe — local, calibrated typed decisions are exactly what we needed for classifying a whole repository file-by-file. Using it today surfaced four gaps that, fixed, would make it dramatically more useful for bulk work like this.

What we hit (macOS arm64, @ruvector/typesafe@0.1.0, Node 24.18)

1. Native ONNX is still a TODO. crates/ruvector-typesafe-ffi/Cargo.toml notes the native-onnx feature's runtime path is unimplemented (the core Engine needs a Box<dyn Embedder> instead of the monomorphized HashEmbedder). So the native addon is hash-only, and real models (bge-small / MiniLM) only run through the WASM tract path.
2. No macOS prebuilt. The npm package ships native/typesafe.linux-x64-gnu.node only, so every Mac falls back to WASM.
3. The documented ONNX entry points fail on WASM. Both of these, as written in the README:
   • createTypesafe({ embedder: { kind: 'onnx', modelDir, manifest } }) → throws the string onnx embedder requires Engine.fromBytes(optionsJson, modelBytes, tokenizerBytes)
   • typesafe decide --embedder onnx --model-dir … --manifest … → prints only error: unknown failure (the thrown value is a string, not an Error, so cli/main.js falls through to the generic branch)
4. Working workaround (for anyone else hitting this):

   const W = require('@ruvector/typesafe/wasm/ruvector_typesafe_wasm.js');
   const m = manifest.models.find((x) => x.name === 'bge-small-en-v1.5');
   const eng = W.Engine.fromBytes(
     JSON.stringify({ embedder: { kind: 'onnx', manifest: JSON.stringify(m), model: m.name }, manifest: JSON.stringify(m) }),
     fs.readFileSync(`${dir}/${m.file}`), fs.readFileSync(`${dir}/${m.tokenizer_file}`));
   eng.decideJson(JSON.stringify({ state, questions }));

   Note manifest must be a single model entry (not the { models: [...] } file), at the top level of the options. With this it classified our probe set correctly (6/6, both bge-small and MiniLM), at ~1.5–2 s per decision on an M3 Max under load.

Why it matters

Bulk classification (e.g. ~1,800 files, several questions, 2–3 embedders for agreement voting) is where typesafe beats an LLM by orders of magnitude — but only at the native p95 ≤ 50 ms target from ADR-006. At WASM speed on Mac it is still usable in parallel workers, but native ONNX would make it ~30–100× faster.

Suggested fixes

• Implement the native-onnx runtime path (Box<dyn Embedder> + OrtEmbedder).
• Publish a darwin-arm64 (and x64) prebuilt.
• Make createTypesafe / the CLI build the WASM ONNX engine from modelDir + manifest via Engine.fromBytes automatically, and accept the manifest file shape.
• Throw Error objects (or wrap strings) so the CLI prints the real message.