perf(parquet): default import chunk size 131072 -> 65536 (#299)

Commit 962520a raised the importer's default chunk from 65536 to 131072 for a
size win, but that predates the FSST + global-dict work. Re-measuring nyc-311
shows the premise reversed: finer chunks now compress better because per-chunk
cost competition and FSST training adapt more locally.

Sweep on nyc-311 (18.49M rows): 131072=2000.00MB, 65536=1934.23MB, 32768=1992.06MB
— 65536 is the sweet spot (below 64k, per-chunk overhead outweighs the gain).
Taxi confirms no regression (40.77 -> 40.63 MB). 65536 also realigns the importer
default with WriteOptions' own default.

Updates the README taxi compression figure to 40.6 MB accordingly.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Author: dfa1

README.md

@@ -42,7 +42,7 @@ same Parquet file (47.6 MB), cascading depth 3, Apple M5:
 | Implementation | Output size | vs Parquet |
 |----------------|-------------|------------|
 | vortex-jni¹    | 47.0 MB     | −1.3%      |
-| **vortex-java** | **40.8 MB** | **−14.2%** |
+| **vortex-java** | **40.6 MB** | **−14.6%** |
 
 ¹ vortex-jni figure measured against 0.78.0 and not re-run (the JNI API has no
 Parquet-import path); the vortex-java and Parquet figures were re-measured 2026-07-24.

parquet/src/main/java/io/github/dfa1/vortex/parquet/ImportOptions.java

@@ -12,7 +12,13 @@ public record ImportOptions(
         WriteOptions writeOptions
 ) {
     public static ImportOptions defaults() {
-        return new ImportOptions(131_072, List.of(), null, WriteOptions.cascading(3));
+        // 65536 physical rows per chunk, matching WriteOptions' default and the Rust reference's
+        // 32k-64k granularity. This was briefly raised to 131072 for a size win (962520ac), but the
+        // FSST + global-dict work (#299) reversed that: finer chunks now compress better on
+        // real-world data (nyc-311 -66 MB, taxi -0.14 MB) as their per-chunk cost competition and
+        // FSST training adapt more locally. 32768 is worse than 65536 — per-chunk overhead outweighs
+        // the finer adaptation below 64k.
+        return new ImportOptions(65_536, List.of(), null, WriteOptions.cascading(3));
     }
 
     /// Restrict import to specific columns. Empty list = all columns.