Test fixtures for ONNX::Native

add.onnx

A minimal ONNX model that computes C = A + B where A, B, and C are all FLOAT[3] tensors. Used by t/03-inference.rakutest to verify an end-to-end round-trip through the shim.

Pinned to ir_version=8 / opset=17 — every ONNX Runtime release since 1.14 supports this combination.

Regenerating

The fixture was produced by a pure-Raku protobuf emitter (no Python onnx package required):

raku t/fixtures/generate-add-onnx.raku

The script writes t/fixtures/add.onnx and prints the byte count. If you change the model, re-run the generator and commit both the script and the new binary.

Why hand-rolled protobuf?

onnx.helper.make_model from the Python onnx package is the standard route. Going through Raku directly avoids adding a Python build-time dep and keeps the fixture reproducible from inside the distribution's own toolchain.

The generator covers only the subset of onnx.proto needed for simple op-graph fixtures (ModelProto, GraphProto, NodeProto, ValueInfoProto, TypeProto.Tensor, TensorShapeProto.Dimension, OperatorSetIdProto). Extending it to emit initializer tensors, attributes, or string tensors would mean implementing more of the protobuf schema — probably worth switching to the Python path at that point.