F32 two-decimal floats compress to 104% of raw (vortex-jni: 22%); F64 skips bitpacking under ALP

[dfa1]

Low-precision floats, such as probabilities rounded to two decimals, compress far worse with vortex-java than with vortex-jni on the same data. For F32 the output is larger than raw.

Data

These are the 1,124 probabilities in the cached jev-1.13.0 answers (jev4j src/test/resources/jev/*.json: yes/no scores, confidences, choice and score distributions).

• Every value has two decimals in [0, 1]. Only 99 distinct values occur; 39% are 0.0 and 10% are 1.0.
• Three values are float artifacts (0.5700000000000001, 0.9299999999999999).
• Shannon entropy is about 4.4 bits/value.

Repro: draw 1M rows at random from those values (new Random(42)), as a single non-nullable column p, written in 65,536-row batches with WriteOptions.defaults() and with vortex-jni 0.86.1 (Arrow Float4/Float8).

Result (1M rows)

	raw	vortex-jni	vortex-java
F64	8.0 MB	915 KB: vortex.alp → fastlanes.bitpacked (~7.3 bits/value)	1.0 MB: vortex.alp → vortex.primitive, no bitpacking (~8.1 bits/value)
F32	4.0 MB	881 KB: vortex.dict → fastlanes.bitpacked (~7.0 bits/value)	4.18 MB: vortex.alp with its i32 integers stored as plain vortex.primitive (~33 bits/value)

Expected

Pick encodings the way Rust does: replicate the vortex-jni choices by diffing the two encoding trees on the same input.

• F32: ALP's encoded i32 child must cascade (narrowing/bitpacking) as the F64 path partly does, and vortex.dict should be a candidate where Rust's float dict scheme wins.
• F64: the ALP integers should be bitpacked, as Rust does.