[dfa1]

Stacked on #624; merge that first. This diff is only the two commits on top.

1. Extension metadata is a typed record. DType.Extension held its metadata as a MemorySegment, a buffer view inside a value type. That needed a hand-written byte-comparing equals/hashCode, a copy off the file mapping so a dtype could outlive its reader, and every consumer re-parsed the bytes. A heap segment from a caller wasn't copied either, so changing its array changed the dtype and its hash.

Rust holds an extension's ExtVTable::Metadata, typed, and converts to bytes only in serialize_metadata/deserialize_metadata. Java now does the same:

• core holds only the types: ExtensionMetadata, TimeUnitMetadata (date, time), TimestampMetadata(TimeUnit, ZoneId) and OpaqueMetadata.
  • ExtensionMetadata is an open interface, since it's an extension point. DType.Extension requires a record, so dtypes compare by value.
  • OpaqueMetadata is Rust's ForeignExtMetadata, kept so a rewritten file loses nothing. The javadoc explains why it isn't null.
• The reader parses:
  • ExtensionMetadataParser handles the well-known ids, straight from the segment with no copy.
  • ExtensionDecoder#parseMetadata handles any other registered id. ReadRegistry.resolveExtensions (formerly validateExtensions) runs it when a file opens.
• The writer serializes:
  • DTypeFbsSerializer writes the built-in records and opaque bytes.
  • ExtensionEncoder#serializeMetadata writes any other record.
  • This is checked at VortexWriter.create, so a dtype whose metadata nothing can serialize fails before any data is written.

Behavior changes: a timestamp zone that ZoneId can't parse now refuses the file at open. Calcite's Temporal read the unit of date and time columns through TimestampDtype.readUnit, which only worked while it read raw bytes; it now uses ExtensionStorage.readUnit.

2. MemorySize moves to core.io, next to IoBounds.

Tests

• CustomExtensionIntegrationTest round-trips example.vector's dimension as a typed VectorMetadata through its own encoder and decoder. It also checks that a writer without the encoder refuses at create.
• ExtensionMetadataParserTest covers the wire-format cases that left ExtensionIdTest.
• ReadRegistryTest covers the decoder hook: nested resolution, and a non-record result wrapped as VortexException.
• The vortex-jni interop tests for date, timestamp, json and the well-known extensions pass unchanged.
• ./mvnw verify is green locally.

Review fixes (b8c567c)

• The writer checks serialized metadata against the reader's 64 KiB cap and refuses an encoder that returns null, both naming the id.
• The reader applies the cap to every id again, including the temporal ones.
• A timestamp zone is kept as the stored string (TimestampMetadata#zone() parses it), as Rust does, so a zone the JDK doesn't know still opens and round-trips byte for byte.
• OpaqueMetadata copies every segment, so a caller's array can't change it later.
• Each extension's metadata is serialized once per writer. A built-in record under a custom id goes through that id's encoder.
• resolveExtensions returns the same node when nothing changed, since it runs per chunk.
• There is one untrusted-id guard (ExtensionMetadataParser.parseId) instead of three.

Examples

• CurrencyExtensionIntegrationTest (c4ccc1a): one java.util.Currency per row over Utf8, next to the vector example. The how-to links both.

🤖 Generated with Claude Code