You are an agent, your current working directory is /app.

You can use the tools available to you to interact with the computer to assist the user in completing tasks.

The workspace under `/app/vendor/onnx` is a frozen slice of a tensor operator used for bounded
attention KV-cache updates. Diagnose and repair the circular-mode indexing bug in the existing
production files:

- `vendor/onnx/onnx/reference/ops/op_tensor_scatter.py`
- `vendor/onnx/onnx/defs/tensor/defs.cc`

Keep the operator's public contract intact. For each batch/prefix coordinate, `write_indices[b]`
selects the sequence start. In `circular` mode, wrap only the selected sequence coordinate modulo
the cache length. Never modulo batch/head prefix coordinates. Preserve untouched cache values,
update ordering, axis normalization, and the existing shape and mode errors. The same sequence-only
rule must be visible in the C++ operator pseudocode.

Run the real source through the supplied workflow:

```sh
cd /app
python3 run_kv_inference.py --input fixtures/kv_cases.json --output /app/output.json
```

The output must be deterministic JSON with `output_schema_version` equal to
`tensor_scatter_inference.v1` and one result for each input case. Do not modify the fixture, runner,
or supporting source files to bypass the repair.