Native-First Design

mifrost is designed so encoding paths produce native BatchEncoding first, with explicit PyG conversion.

Why this helps:

• Keeps the fast path in native columnar buffers.
• Avoids immediate graph-object materialization when not needed.
• Makes conversion an explicit boundary (as_pyg, encode_pyg, encode_batch_pyg, flush_pyg).
• Allows model-boundary tensor access via as_hetero() / as_homo() facades without creating PyG objects.

Recommended pattern:

1. Encode to native BatchEncoding.
2. Perform any native batching/serialization operations.
3. Use as_hetero() / as_homo() for tensor-first model code when possible.
4. Convert to PyG only for APIs that require real Data / HeteroData.