feat(memory): JVM/Android Storage bindings — SegmentStorage (FFM), MappedFileStorage, DirectBufferStorage, MappedBufferStorage (SKEEP-003 P2, S1.1b)

[michalharakal]

Summary

SKEEP-003 slice S1.1b (milestone M1 #1002, PRD M1-F1): the JVM and Android bindings of the OffHeap and Mapped storage kinds (§4.8.1 / §4.8.2) — where bytes live is a policy decision; these are the concrete kinds the planner and scopes pick from.

• SegmentStorage : Storage.OffHeap (jvmMain, FFM): a MemorySegment in an Arena. allocate(bytes, scope, arena?) owns a private Arena.ofShared() and frees it on close(), or takes the scope's arena (the Forward bump-slab pattern of [S1.2] P2: Scope Model/Forward/Ambient — JVM Arena bump allocator, reset(), retain(), escape check; ExecutionContext opt-in #1021); borrow(segment) / borrow(FloatArray) via MemorySegment.ofArray — zero-copy, never freed by us (the GGUF DEQUANTIZE_TO_FP32 over-allocates: 1.1B Q4_K_M needs >12 GB heap transiently (~4.4 GB legit) #782 pattern); slice() = asSlice alias; segment() is the once-per-call accessor and throws StorageClosedException after close.
• MappedFileStorage : Storage.Mapped (jvmMain): FileChannel.map(READ_ONLY) into an Arena.ofShared(); closing unmaps; path + fileOffset kept for describe(); slices carry their file offset.
• DirectBufferStorage : Storage.OffHeap and MappedBufferStorage : Storage.Mapped (jvmAndroidMain): direct ByteBuffer outside the ART heap (createRandomAccessSource returns null on Android: full-file heap load OOMs on a 138 MiB GGUF (working ~40-line fix included) #922) and MappedByteBuffer (SKEEP-002 / Android tensor storage is ART-heap-bound: no off-heap / mmap path outside jvmMain caps practical model size (SKEEP candidate) #921); little-endian duplicates handed out per call; close() marks the storage dead.
• Allocation / Free trace events for all kinds; MemoryDomain.HOST_OFFHEAP / MMAP_FILE.
• Android note: ByteBuffer.position/limit return Buffer under the Android (Java 8) signatures, so the slice path keeps a ByteBuffer-typed local — compileAndroidMain is part of the gate's assemble.
• JvmStorageTest: arena ownership vs caller arena, alias over the same bytes, borrowed array zero-copy, mapped file read + unmap, direct and mapped buffers, distinct ids across kinds.
• BCV: lang-core jvm dump regenerated (additions only). Nothing consumes these yet (Scope [S1.2] P2: Scope Model/Forward/Ambient — JVM Arena bump allocator, reset(), retain(), escape check; ExecutionContext opt-in #1021, façades [S1.4a] P2: dense TensorData implementations become façades over TensorView (bit-identical) #1023/[S1.4b] P2: packed TensorData implementations (Q4_0…Q8_0, Q4_K/Q5_K/Q6_K, ternary, TurboQuant) become façades over TensorView #1024, loader config [S2.4] P5: IO as a pipeline — quantPolicy × staging = MAPPED|HEAP in one loader; suspend RandomAccessSource in io-core #1037/[S2.5] P5: Android mmap as a loader configuration (SKEEP-002, #921, #922) on Storage.Mapped #1038).

Stacked on #1061 (TraceSink + Storage, after #1062 merged into that branch); retarget to develop after it merges.

Test plan

Full local gate (scripts/pr-gate.sh, JDK 25); results in the first comment. Targeted: lang-core sk.ainet.lang.memory.* 44/44 (jvmTest).

Closes #1019

🤖 Generated with Claude Code

[michalharakal]

Local gate scripts/pr-gate.sh (JDK 25) on 4f61f29 (rebased onto #1061 after #1062 merged into it): all legs passed — jvmTest · apiCheck · verifyNpmPins jsTest wasmJsTest wasmWasiTest · linuxX64Test · assemble (incl. compileAndroidMain) · :skainet-test:skainet-test-java:test.

Targeted: lang-core sk.ainet.lang.memory.* 44/44 (jvmTest).