recipes

Recipes — content-addressed pipeline caching over the type system.

Decomposed pure functions over CGType values become memoized, garbage-collected pipeline stages, keyed by merkle hash in a managed scratch. The package is three layers, bottom up, each ignorant of the one above:

• [.scratch][] — the store. A garbage-collected, content-addressed store of directory trees: objects, names, pins, and locks. It knows nothing about types.
• [.values][] — where a typed value meets the store: how it becomes an object tree, how it hashes, how a call's arguments hash, and the workspace a call writes into.
• [.calls][] — memoized calls and keyed state: pure_function and ImpureState, the surface a recipe author touches.
• [.progress][] — what a computation is doing right now, as a tree of steps every memoized call, map and pmap reports into; cg.step for a body's own, cg.print_progress to watch from a terminal.
• [.splats][] — the one recipe library shipped with the package: 3D Gaussian splats from posed captures, trained on the GPU under cg.lock("gpu").

testing ships the cg_scratch pytest fixture. This __init__ is documentation-only. The root exports scratch and splats as folders (cg.scratch.configure(), cg.splats.train()) and imports the values / calls leaves directly (cg.pure_function, cg.ImpureState, cg.value_hash).