fix(reader): replace slab with per-chunk Arena to fix OOB on large layouts

SlicingAllocator on a fixed 4 MB slab failed two ways:
1. Whole-file dict layouts (e.g. symbol column) needed a single
   20 MB allocation → immediate OOB on first chunk.
2. Slab reset between chunks (prior fix) made previous chunk's
   MemorySegments point to overwritten memory.

Replace with Arena.ofConfined() per chunk: previous arena is closed
before decoding the next chunk; last arena is closed on natural
exhaustion or explicit close(). Data is valid until the next
hasNext() call — the intended API contract.

Author: dfa1

README.md

@@ -11,7 +11,7 @@ Pure-Java reader/writer for the [Vortex](https://github.com/spiraldb/vortex) col
 - Pure-Java reader for primitive, sequence, ALP, dict, FSST (stable)
 - Local (mmap) or Remote (HTTPS, single read of last 65K) (stable)
 - Writer: in progress
-- Benchmark vs Rust+JNI: in progress
+- Benchmark vs Rust+JNI: Java beats JNI 1.5×–6.5× across read/write workloads (see Benchmarks)
 - Full encoding coverage: in progress
 - Vectorized decode paths (Panama Vector API): planned
 - Iceberg/Spark/Flink integration: not available yet
@@ -81,6 +81,32 @@ file occupies. Decoding reads bytes directly from that view with no copies, no i
 Arrow format, and no boundary crossings. The JIT sees the full decode path as ordinary Java
 bytecode.
 
+## Benchmarks
+
+JMH throughput (ops/s = full-file scans per second). Higher is better.
+
+**Environment:** Apple M5, OpenJDK 27-jep401ea3 (Valhalla EA), 3 warmup × 3 s, 5 measurement × 5 s, fork 1.
+
+### OHLC read — 10 M rows, 58.9 MB (Rust-written file, single-column projection)
+
+| Benchmark      | Java (ops/s)     | JNI/Rust (ops/s) | Java speedup |
+|----------------|------------------|------------------|--------------|
+| close (F64/ALP)| 76.4 ± 1.6       | 50.8 ± 2.2       | **1.5×**     |
+| volume (I64)   | 128.6 ± 1.5      | 52.3 ± 1.1       | **2.5×**     |
+| symbol (varbin)| 63.4 ± 21.6      | 9.8 ± 1.4        | **6.5×**     |
+
+### OHLC write — 10 M rows
+
+| Benchmark | Java (ops/s) | JNI/Rust (ops/s) | Java speedup |
+|-----------|--------------|------------------|--------------|
+| write     | 4.6 ± 0.6    | 0.7 ± 0.0        | **6.5×**     |
+
+### Big-file scan — 100 M rows × 4 I64 columns, ~3 GB (Rust-written file, all columns)
+
+| Benchmark | Java (ops/s) | JNI/Rust (ops/s) | Java speedup |
+|-----------|--------------|------------------|--------------|
+| scan      | 20.3 ± 1.0   | 5.7 ± 0.2        | **3.6×**     |
+
 ## Design principles
 
 - Zero-copy everywhere

reader/src/main/java/io/github/dfa1/vortex/io/VortexHttpReader.java

@@ -174,7 +174,7 @@ public MemorySegment slice(long offset, long length) {
 
 	@Override
 	public ScanIterator scan(ScanOptions options) {
-		return new ScanIterator(this, options, arena);
+		return new ScanIterator(this, options);
 	}
 
 	@Override

reader/src/main/java/io/github/dfa1/vortex/io/VortexReader.java

@@ -144,7 +144,7 @@ public EncodingRegistry registry() {
 
 	@Override
 	public ScanIterator scan(ScanOptions options) {
-		return new ScanIterator(this, options, this.arena);
+		return new ScanIterator(this, options);
 	}
 
 	/// Zero-copy slice of the memory-mapped file.

reader/src/main/java/io/github/dfa1/vortex/scan/ScanIterator.java

@@ -47,13 +47,9 @@ public final class ScanIterator implements AutoCloseable {
 	private static final ValueLayout.OfInt   LE_INT   = ValueLayout.JAVA_INT_UNALIGNED.withOrder(ByteOrder.LITTLE_ENDIAN);
 	private static final ValueLayout.OfLong  LE_LONG  = ValueLayout.JAVA_LONG_UNALIGNED.withOrder(ByteOrder.LITTLE_ENDIAN);
 
-	private static final long SLAB_SIZE = 4 * 1024 * 1024L;
-
 	private final VortexHandle file;
 	private final ScanOptions options;
-	private final Arena arena;
-	private MemorySegment chunkSlab;
-	private SegmentAllocator chunkAlloc;
+	private Arena chunkArena;
 
 	private List<ChunkSpec> chunks;
 	private Map<String, DType> columnDtypes;
@@ -63,10 +59,9 @@ public final class ScanIterator implements AutoCloseable {
 	private long rowsReturned;
 	private ScanResult current;
 
-	public ScanIterator(VortexHandle file, ScanOptions options, Arena arena) {
+	public ScanIterator(VortexHandle file, ScanOptions options) {
 		this.file = file;
 		this.options = options;
-		this.arena = arena;
 	}
 
 	private static void collectFlats(Layout layout, List<Layout> out) {
@@ -134,11 +129,15 @@ public boolean hasNext() {
 				continue;
 			}
 
-			chunkAlloc = SegmentAllocator.slicingAllocator(chunkSlab);
+			if (chunkArena != null) {
+				chunkArena.close();
+			}
+			chunkArena = Arena.ofConfined();
 			current = new ScanResult(chunk.rowCount(), buildColumnMap(chunk));
 			rowsReturned += chunk.rowCount();
 			return true;
 		}
+		close();
 		return false;
 	}
 
@@ -153,6 +152,10 @@ public ScanResult next() {
 
 	@Override
 	public void close() {
+		if (chunkArena != null) {
+			chunkArena.close();
+			chunkArena = null;
+		}
 	}
 
 	// ── Flat segment decoding ─────────────────────────────────────────────────
@@ -187,8 +190,6 @@ private void initialize() {
 		chunks = buildChunks(columnFlats);
 		projectedNames = List.copyOf(columnDtypes.keySet());
 		projectedDtypes = List.copyOf(columnDtypes.values());
-		chunkSlab = arena.allocate(SLAB_SIZE, 64);
-		chunkAlloc = SegmentAllocator.slicingAllocator(chunkSlab);
 	}
 
 	// ── Zone-map pruning ──────────────────────────────────────────────────────
@@ -200,20 +201,20 @@ private Map<String, Array> buildColumnMap(ChunkSpec chunk) {
 		Layout[] layouts = chunk.columnLayouts();
 		int n = projectedNames.size();
 		if (n == 1) {
-			Array arr = decodeLayout(layouts[0], projectedDtypes.get(0), chunkAlloc);
+			Array arr = decodeLayout(layouts[0], projectedDtypes.get(0), chunkArena);
 			if (arr instanceof StructArray sa) {
 				return expandStruct(sa);
 			}
 			return Map.of(projectedNames.get(0), arr);
 		}
 		if (n == 2) {
 			return Map.of(
-					projectedNames.get(0), decodeLayout(layouts[0], projectedDtypes.get(0), chunkAlloc),
-					projectedNames.get(1), decodeLayout(layouts[1], projectedDtypes.get(1), chunkAlloc));
+					projectedNames.get(0), decodeLayout(layouts[0], projectedDtypes.get(0), chunkArena),
+					projectedNames.get(1), decodeLayout(layouts[1], projectedDtypes.get(1), chunkArena));
 		}
 		var scratch = new LinkedHashMap<String, Array>(n);
 		for (int i = 0; i < n; i++) {
-			scratch.put(projectedNames.get(i), decodeLayout(layouts[i], projectedDtypes.get(i), chunkAlloc));
+			scratch.put(projectedNames.get(i), decodeLayout(layouts[i], projectedDtypes.get(i), chunkArena));
 		}
 		return Map.copyOf(scratch);
 	}