Pluggable extension types: register custom logical types on the read/write registries #530

Rust lets users add a logical type by implementing ExtVTable (id(), validate_dtype(), metadata/scalar handling) and registering it on the session (session.dtypes().register(Vector)); see Adding a new logical type.

Current state (not supported)

• ExtensionId is a closed enum (VORTEX_DATE/TIME/TIMESTAMP/UUID), so a third-party id cannot even be named.
• Read side: ExtensionDecoders are singletons dispatched by a case in Chunk.as(), not registry-managed.
• Write side: WriteRegistry.Builder#register(ExtensionEncoder) exists but can only re-implement the four built-ins.
• CLAUDE.md says DType is pluggable only via Extension, but today that only works for wire-known ids, and docs/reference.md documents the gap.

Proposal

• Make ExtensionId a sealed WellKnown/Custom type, like EncodingId/LayoutId.
• Register ExtensionDecoder on ReadRegistry.Builder (built-ins via registerDefaults()), replacing the switch in Chunk.as().
• Add a storage-dtype validation hook (Rust's validate_dtype) on both sides.
• Unknown extension ids: decode as the storage array (Rust default), with an opt-in/out policy to decide.
• Interop test: a Rust-written custom extension (e.g. a Vector over FixedSizeList<f32>) read by Java, and vice versa.
• Update docs/reference.md, CLAUDE.md and the CHANGELOG in the same change.

Out of scope: compute functions on custom types (Rust vortex-tensor-style scalar functions).