Architecture Status

Current Phase

Tensor4all.jl is in a review-first skeleton phase.

The package now exposes real metadata-level types and integration boundaries for review, while backend numerics remain intentionally stubbed.

Layer	Responsibility	Current status
Core	lazy backend loading, common errors, `Index`, `Tensor`	metadata behavior implemented
TreeTN	`TreeTensorNetwork`, chain aliases, runtime topology predicates	metadata behavior implemented
Quantics reuse	adopted `QuanticsGrids.jl` grid and coordinate-conversion surface	curated re-export implemented
Quantics local layer	transform descriptors and QTCI placeholders	metadata/stub behavior implemented
Extensions	ITensors and HDF5 compatibility glue	extension-only stubs implemented
BubbleTeaCI	`TTFunction` and high-level function workflows	intentionally out of scope here

Metadata constructors, inspection helpers, and topology predicates work.
Backend-backed operations such as contraction, dense materialization, and transforms deliberately throw SkeletonNotImplemented.
Importing Tensor4all.jl does not require a compiled tensor4all-rs backend.

tensor4all-rs owns kernels, storage, and numerically heavy backend behavior.
Tensor4all.jl owns Julia-side wrappers, TreeTN-general abstractions, local quantics transforms, QTCI placeholders, and compatibility extensions.
QuanticsGrids.jl owns quantics grid semantics and coordinate conversion.
Tensor4all.jl re-exports a curated QuanticsGrids.jl surface for usability, but that re-export does not change ownership.
BubbleTeaCI remains the home of TTFunction and high-level function workflows, and should consume lower-level functionality from Tensor4all.jl and adopted dependencies instead of duplicating it.

Should Index and Tensor gain explicit backend-handle behavior beyond the current nullable-handle skeleton fields before backend integration starts?
Is the current curated QuanticsGrids.jl re-export scope broad enough for the first downstream consumers?
Is the downstream BubbleTeaCI contract explicit enough before migration begins?
Are the current public names good enough to freeze for backend implementation?