Testing conventions

Layout

Tests mirror the source tree, one directory per sub-module, aggregated by test/runtests.jl:

test/
  Elliptic/  Core/  Elasticity/  Cracks/  Conductivity/
  LayeredSpheres/  LayeredSpheroids/  Schemes/  Viscoelasticity/
  regression/          # cross-cutting cases spanning the dispatch surface
  runtests.jl

runtests.jl does three things before any include, and all three matter:

it imports DECUHR, Integrals and NonlinearSolve so the package extensions activate — several tests cross-validate method = :decuhr against the residue and nested-QuadGK backends, and exercise the SciML self-consistent solvers. Without those imports the corresponding methods do not exist and the tests silently cover less than they appear to;
it calls Random.seed! once, so a CI failure is reproducible locally instead of depending on the draw;
it wraps everything in a single top-level @testset "MeanFieldHom".

What a good test looks like here

Every new public function ships with at least a smoke test — but a smoke test is a floor, not a target. The tests that actually catch bugs in this package are of three kinds:

Closed-form oracles. Where an analytical result exists, assert against it at tight tolerance; rtol = 1e-13 is routinely achievable. Example: the penny-crack conduction COD $b = 2/(\pi^{2}K)$ in test/Cracks/test_thermal.jl.
Cross-backend agreement. The same quantity computed two independent ways must agree — :residues vs :decuhr; quadrature vs the BigFloat monomial series (test/LayeredSpheroids/test_coupling.jl); the general series solution vs the perfect-interface closed form (test/LayeredSpheroids/test_conductivity.jl).
Invariants. Identities that hold whatever the numbers: $\sum_i I_i^{\boldsymbol{A}} = 1$, the Eshelby identities, or the contribution invariant $\boldsymbol{N}_K = \boldsymbol{B}_\Omega - \boldsymbol{K}_0\cdot\boldsymbol{A}_\Omega$ for a layered particle.

Test an asymmetric case, not only the symmetric one

Degenerate cases hide convention errors, because competing conventions tend to agree there. A penny crack has $\eta = 1$, and at $\eta = 1$ the normalizations of the crack compliance $\mathbb{H}$ used by MeanFieldHom, by Echoes and by the literature all coincide (see Crack opening displacement, section Conventions). A suite covering only the penny therefore cannot detect a wrong $\eta$-dependence.

This is not hypothetical. As of 2026-07 the elastic $\mathbb{H}$ is covered by a smoke test plus penny and ribbon cases only, so the elliptic $\eta \ne 1$ behaviour of the $\boldsymbol{B} \to \mathbb{H}$ factor is pinned by no test at all. The same trap applies to spheres among spheroids and to isotropy among symmetry classes: always add one asymmetric case.

Number types are part of the contract

The package is generic over the scalar type, and that genericity is advertised in the dispatch tables of the theory pages — so it is tested explicitly:

type	role
`Float64`	the default path
`ForwardDiff.Dual`	every algorithm advertised as AD-safe must be exercised through `derivative`, not merely called with a `Dual`
`BigFloat`	precision oracle, notably for the spheroid coupling matrices
symbolic (`SymPy`, `Symbolics`)	closed-form paths only

When adding an algorithm, state its AD compatibility in the docstring and back that claim with a test.

Coverage

Run coverage through .github/scripts/coverage.jl rather than the stock julia-processcoverage action.

The stock processor misattributes lines inside generated functions and multi-line expressions — observed reporting ~59 % where the true figure is ~95 %. Trust the script's numbers over the action's.