Improved saved world startup

- Startup holds generation admission through recovery and runtime attachment
- Prepared startup state removes the second recovery scan before generator readiness
- Recovery batches allocated chunks by region with reusable snapshot buffers
- Region decoding uses up to four workers within processor and heap limits
- Oversized regions and external chunks use bounded direct reads
- Status parsing validates chunk coordinates without constructing the full NBT tree
- Region snapshots reject changed files, allocation headers and truncated payloads
- Native zlib decompression uses jlibdeflate with a Java inflater fallback
- Chunk readers reuse region handles and decompression buffers during recovery and boundary capture
- Semantic shards reuse validated summaries with invalidation on mutation
- Semantic key validation avoids regular expression allocation
- Saved biome indexes use buffered sequential reads with reusable CRC state
- Pack inspection reads validated manifest and pack state without opening terrain indexes
- Datapack compilation reuses pack inspection within each operation
- Kernel staging avoids duplicate pack hashing when no activation is pending
- Pack copying and directory synchronization use up to four available processors
- Boundary publication batches directory synchronization around durable catalog publication
- Configured world reconciliation reuses ready runtimes after identity and saved-state checks
- Explicit world loads retain the full validation path
- Regression tests cover startup admission, recovery, malformed data, resource limits, publication and reconciliation

core/build.gradle

@@ -150,6 +150,7 @@ dependencies {
     slim(libs.commons.lang3)
     slim(libs.oshi)
     slim(libs.lz4)
+    slim('com.fulcrumgenomics:jlibdeflate:0.2.0')
     slim(libs.fastutil)
     slim(libs.zip) {
         exclude(group: 'org.slf4j', module: 'slf4j-api')