fix(writer): exclude I8/I16 from global dict — reader can't decode narrow-int dict

[dfa1]

Fixes the write/read incompatibility surfaced by mutation coverage in #99.

The bug

The writer's global dict admitted I8/U8/I16/U16 columns, but the reader's lazy dict decode (ScanIterator.buildLazyDictPrimitive) only supports I32/I64/F32/F64. A low-cardinality I16 column therefore wrote a vortex.dict the reader rejected on read-back:

VortexException: vortex.dict: layout: unsupported ptype for lazy dict: I16

i.e. the writer produced an unreadable file.

Cross-checked against Rust (ground truth)

New RustWritesJavaReadsIntegrationTest#jniWriter_javaReader_lowCardinalityI16: the Rust/JNI writer encodes a 10k-row, 3-distinct I16 column and the Java reader reads it back exactly — no dict error. So Rust does not dict-encode narrow ints, and the Java reader is already Rust-conformant. The Java writer was the outlier.

Fix

Exclude I8/U8/I16/U16 from isDictCandidate (alongside F16/F32):

• a U8/U16 code is no smaller than the value → dict gives no real benefit;
• matches the Rust compressor;
• only ever produced files the reader couldn't read → not a regression.

A low-card I16 column now encodes via the cascade and round-trips cleanly (GlobalDictPrimitiveTest#lowCardinality_i16_notDicted_roundTrips), and VortexWriterDictDecisionTest pins the new exclusion.

Verify

./mvnw -pl writer -am test
./mvnw verify -pl integration -am -Dit.test="RustWritesJavaReadsIntegrationTest#jniWriter_javaReader_lowCardinalityI16"

🤖 Generated with Claude Code