feat: replace SSTables with Parquet, add predicate partitioning and tiered cache (Phase 2)

Replace custom SSTable binary format with Apache Parquet columnar storage,
introduce vertical partitioning by predicate, and add a three-tier cache
(Caffeine heap -> local disk LRU -> S3).

Storage redesign:
- Parquet files on S3 with ZSTD compression and dictionary encoding
- Predicate-based partitioning (data/predicates/{id}/) eliminates
  predicate column from files, tightening column statistics
- Three sort orders per partition (SOC, OSC, CSO) for optimal query
  performance regardless of access pattern
- Single MemTable in SPOC order, partitioned on flush
- JSON catalog with per-file column statistics for catalog-level pruning

Cache system:
- L1: Caffeine heap cache (configurable, default 256 MB)
- L2: Local disk LRU cache (configurable, default 10 GB)
- L3: S3 source of truth
- Write-through on flush avoids cold reads

Compaction:
- L0->L1 merge when epoch count >= 8 per predicate
- L1->L2 merge when epoch count >= 4 per predicate
- Tombstone suppression at highest level

Hadoop dependency elimination:
- Zero Hadoop JARs in dependency tree
- PlainParquetConfiguration + custom SimpleCodecFactory bypass all
  Hadoop runtime paths
- 14 minimal stub classes in org.apache.hadoop.* satisfy parquet-hadoop
  JVM class loading requirements

Deleted: SSTable, SSTableWriter, Manifest (replaced by Parquet + Catalog)
All 529 tests pass.

Author: odysa

core/sail/s3/pom.xml

@@ -52,6 +52,26 @@
 			<groupId>com.fasterxml.jackson.core</groupId>
 			<artifactId>jackson-databind</artifactId>
 		</dependency>
+		<dependency>
+			<groupId>org.apache.parquet</groupId>
+			<artifactId>parquet-hadoop</artifactId>
+			<version>1.15.2</version>
+			<exclusions>
+				<exclusion>
+					<groupId>javax.annotation</groupId>
+					<artifactId>javax.annotation-api</artifactId>
+				</exclusion>
+			</exclusions>
+		</dependency>
+		<!-- No Hadoop jars needed. Parquet-hadoop references a few Hadoop classes
+		     in method signatures and superclass declarations, but we bypass all
+		     Hadoop runtime paths via PlainParquetConfiguration + SimpleCodecFactory.
+		     Minimal stub classes in org.apache.hadoop.* satisfy JVM class loading. -->
+		<dependency>
+			<groupId>com.github.ben-manes.caffeine</groupId>
+			<artifactId>caffeine</artifactId>
+			<version>3.1.8</version>
+		</dependency>
 		<dependency>
 			<groupId>${project.groupId}</groupId>
 			<artifactId>rdf4j-sail-testsuite</artifactId>

core/sail/s3/src/main/java/org/apache/hadoop/conf/Configuration.java

@@ -0,0 +1,15 @@
+/*
+ * Minimal stub for org.apache.hadoop.conf.Configuration.
+ *
+ * Parquet-hadoop references this class in abstract method signatures
+ * (WriteSupport.init, ParquetWriter.Builder.getWriteSupport). Our code
+ * overrides the ParquetConfiguration variants instead, so this class is
+ * never instantiated or used at runtime. It exists only to satisfy the
+ * JVM class loader.
+ */
+package org.apache.hadoop.conf;
+
+public class Configuration {
+	public Configuration() {
+	}
+}

core/sail/s3/src/main/java/org/apache/hadoop/fs/FileStatus.java

@@ -0,0 +1,8 @@
+/*
+ * Minimal stub — satisfies JVM class loading for parquet-hadoop.
+ * Never instantiated at runtime.
+ */
+package org.apache.hadoop.fs;
+
+public class FileStatus {
+}

core/sail/s3/src/main/java/org/apache/hadoop/fs/FileSystem.java

@@ -0,0 +1,8 @@
+/*
+ * Minimal stub — satisfies JVM class loading for parquet-hadoop.
+ * Never instantiated at runtime.
+ */
+package org.apache.hadoop.fs;
+
+public abstract class FileSystem {
+}

core/sail/s3/src/main/java/org/apache/hadoop/fs/Path.java

@@ -0,0 +1,8 @@
+/*
+ * Minimal stub — satisfies JVM class loading for parquet-hadoop.
+ * Never instantiated at runtime.
+ */
+package org.apache.hadoop.fs;
+
+public class Path {
+}

core/sail/s3/src/main/java/org/apache/hadoop/fs/PathFilter.java

@@ -0,0 +1,8 @@
+/*
+ * Minimal stub — satisfies JVM class loading for parquet-hadoop.
+ * Never instantiated at runtime.
+ */
+package org.apache.hadoop.fs;
+
+public interface PathFilter {
+}

core/sail/s3/src/main/java/org/apache/hadoop/mapred/JobConf.java

@@ -0,0 +1,10 @@
+/*
+ * Minimal stub — satisfies JVM class loading for parquet-hadoop.
+ * Never instantiated at runtime.
+ */
+package org.apache.hadoop.mapred;
+
+import org.apache.hadoop.conf.Configuration;
+
+public class JobConf extends Configuration {
+}

core/sail/s3/src/main/java/org/apache/hadoop/mapreduce/InputFormat.java

@@ -0,0 +1,8 @@
+/*
+ * Minimal stub — satisfies JVM class loading for parquet-hadoop.
+ * Never instantiated at runtime.
+ */
+package org.apache.hadoop.mapreduce;
+
+public abstract class InputFormat<K, V> {
+}

core/sail/s3/src/main/java/org/apache/hadoop/mapreduce/InputSplit.java

@@ -0,0 +1,8 @@
+/*
+ * Minimal stub — satisfies JVM class loading for parquet-hadoop.
+ * Never instantiated at runtime.
+ */
+package org.apache.hadoop.mapreduce;
+
+public abstract class InputSplit {
+}

core/sail/s3/src/main/java/org/apache/hadoop/mapreduce/Job.java

@@ -0,0 +1,8 @@
+/*
+ * Minimal stub — satisfies JVM class loading for parquet-hadoop.
+ * Never instantiated at runtime.
+ */
+package org.apache.hadoop.mapreduce;
+
+public class Job {
+}