EarthSciDiscretizations

Authoritative discretization rule files — and the conformance machinery that keeps them honest — for the EarthSci model stack.

This repository has three roles:

1. A catalog of declarative discretization rule JSON files under discretizations/ (finite-difference and finite-volume stencils, reconstructions, flux forms, and boundary conditions), validated against the EarthSciSerialization (ESS) §7 discretization schema and executed by the ESS rule engine via the canonical discretize → ArrayOp → eval pipeline. The package side contributes structural lowering between rule forms (src/stencil_lowering.jl) and thin coefficient-evaluation passthroughs (src/rule_eval.jl). This catalog is the repository's primary, durable role.
2. Grid runtimes and conformance machinery. Grid constructors for cartesian, lat-lon, vertical, Arakawa-staggered, cubed-sphere, MPAS, and DUO topologies (src/grids/), mirrored by thin Python, Rust, and TypeScript bindings (python/, rust/, typescript/), plus the cross-binding conformance suite: golden fixtures under tests/conformance/, the dynamic rule walker (test/walk_esd_tests.jl), and its machine-readable per-rule skip ledger (test/junit-esd.xml). One catalog, four bindings, one canonical pipeline — conformance-tested.
3. Retired reference operators. Hand-coded finite-volume operators on cubed-sphere grids (transport_2d, flux_1d, PPM transport/reconstruction, ghost-cell handling) formerly lived under src/operators/ as oracles while their math was ported into catalog rules. All have now been retired; src/operators/ no longer exists (see the retirement log in src/EarthSciDiscretizations.jl). fv_laplacian was retired in esd-ecq; its math now lives entirely in discretizations/finite_difference/covariant_laplacian_cubed_sphere.json. New math should land as rules, not operators.

Per the single-pathway rule in AGENTS.md, ESD is a discretization catalog over ESS, not a parallel runtime: no binding carries a rule evaluator, and every simulation flows through the ESS discretize → ArrayOp → eval pipeline.

See docs/REPO_LAYOUT.md for the repository convention and discretizations/README.md for how to add a rule file.

Development

julia --project=. -e 'using Pkg; Pkg.test()'