New Arnold Pipeline

Skill: Create a New Arnold Native-First Pipeline

1. Copy arnold_pipelines/_template/ to arnold_pipelines/<your_pipeline>/.
2. Edit __init__.py: set name, description, capabilities, keep driver=("native", "project+validate") and supported_modes=("native",), and ensure build_pipeline() compiles the native program and returns a projected Pipeline with a non-null native_program.
3. Edit pipelines.py: replace the skeleton @phase functions with real logic. Use @pipeline, @phase, @decision, parallel, compile_pipeline, and project_graph from the public arnold.pipeline authoring surface.
4. Run arnold pipelines check <your_pipeline> to validate the package against the native-first authoring contract (metadata, driver, native program, and graph projection).
5. Add tests that compile the native program, project the graph, and assert the returned Pipeline carries a non-null native_program.

Dispatch Substrate Note

The native_program compiled by build_pipeline() is a dispatch substrate — it describes how the runtime executes the pipeline, but it does not define the final visible compositional semantics. Panel synthesis, join delegation, parallel merge strategy, subpipeline ownership, and Capsule projection are deferred to later Megaplan layers.

Do not add _legacy.py, graph fallback builders, compatibility namespaces, shim packages, or temporary wrapper modules. The template and this workflow are native-first only. Generated manifests and catalogs are outputs of the authoring path, not editable source.