File IO

File I/O

src/io/ — public always; implementation behind feature hdf5.

Function	Behavior
`io::read(path)`	Load `.nir` → `NirGraph`
`io::read_version(path)`	Strict `/version` (errors if absent)
`io::write(path, graph)`	Default options
`io::write_with(path, graph, opts)`	Compression, version override, validate flag

Without feature: all return NirError::Unimplemented(...).

Field	Default	Notes
`compression`	`Some(4)`	Deflate 0..=9 for numeric arrays; None = uncompressed
`version`	None	Overrides `graph.version`; else `DEFAULT_NIR_VERSION` = `"1.0.8"`
`validate`	`true`	Runs structure + wire checks before truncating file

Scalars, empty arrays, and string datasets are never compressed (HDF5 limits / no benefit).

Permissive: does not auto-validate_structure (callers decide)
Optional Python-default fill for missing fields
Graph-level round-trip fidelity, not byte-identical to h5py
Safety: external links rejected; nested graph budget; per-dataset size guard (see code comments / #21 for cumulative budget)

Pre-checks names, compression range, string NULs, usize extents before File::create so a bad graph does not wipe an existing model.