Kill the naked int/long size: ZstdByteSize value type

[dfa1]

Summary

• Adds ZstdByteSize, a record wrapping a byte size/count as a validated non-negative long, with toIntExact() for callers that must narrow to an int because the result ends up in a byte[] (throws ArithmeticException past Integer.MAX_VALUE, matching Math.toIntExact).
• Replaces every naked int/long size parameter with it: Zstd.decompress, Zstd.compressBound, Zstd.estimateCompressDictSize/estimateDecompressDictSize, ZstdDictionary.train/trainCover/trainFastCover/finalizeFrom, ZstdOutputStream.withPledgedSize.
• Centralizes interpretation of zstd's raw frame-content-size results as ZstdByteSize factories rather than scattered sentinel checks:
  • ZstdByteSize.fromFrameContentSize(long) turns the ZSTD_getFrameContentSize/ZSTD_findDecompressedSize/ZSTD_decompressBound sentinels — and any other negative reading (an unsigned value ≥ 2^63) — into a ZstdException instead of a bogus negative size or the wrong exception type. Used by Zstd.decompress(byte[]), Zstd.decompressedSize(MemorySegment), and ZstdFrame.decompressedSize/decompressedBound, which now return ZstdByteSize instead of unwrapping immediately back to long.
  • ZstdByteSize.fromFrameHeaderContentSize(long) backs ZstdFrameHeader.contentSize(), now Optional<ZstdByteSize> instead of OptionalLong.
  • The CONTENTSIZE_UNKNOWN/CONTENTSIZE_ERROR sentinel constants move from Zstd to become a private implementation detail of ZstdByteSize.
• ZstdFrame's other size-returning methods (compressedSize, decompressionMargin, headerSize) are unaffected — they never construct a ZstdByteSize, no sentinel semantics involved.
• Breaking change (pre-1.0, no compatibility shims). CHANGELOG updated under [Unreleased].

Closes #96.

Design notes

• One long-backed type, not two split by width — narrowing happens at the four array-bounded call sites via toIntExact(); native/streaming totals pass long straight through. See discussion on Consider a size value type for byte-size parameters #96.
• A real regression was caught and fixed during review: a frame declaring a content size with the sign bit set (unsigned ≥ 2^63) was falling through both named sentinel checks and throwing the wrong exception type (IllegalArgumentException instead of the documented ZstdException). Covered by a new regression test (ZstdTest.UntrustedInput.rejectsAFrameDeclaringAContentSizeWithTheSignBitSet).

Test plan

• ./mvnw -pl zstd,integration-tests,benchmark -am clean verify — build success, all unit + integration tests pass
• ./mvnw -pl zstd validate — checkstyle clean
• ./mvnw -pl zstd javadoc:javadoc — zero output (failOnError/failOnWarnings)
• Added ZstdByteSizeTest covering construction, negative rejection, toIntExact() in-range/overflow
• Manually reproduced and verified the sign-bit-set regression fix against a hand-crafted frame
• Multiple independent review passes focused on native-call boundary correctness, exception-type consistency, and Javadoc/call-site accuracy — no outstanding issues

🤖 Generated with Claude Code

[dfa1]

maybe it should be called ZstdSize?

[dfa1]

maybe we should also expose a better way to build 8K or 1M constants? this could be useful for tests internally but also for end users to define the buffer size for copies, etc

[dfa1]

Re: ZstdSize naming — kept ZstdByteSize. It matches the house convention of descriptive multi-word names (ZstdCompressionLevel, ZstdDictionaryId, ZstdFrameHeader) — "Size" alone felt more ambiguous (could read as a count, a dimension) where "ByteSize" says exactly what's inside.

Re: KB/MB constants — added ZstdByteSize.ofKiB(long) / ofMiB(long) (using Math.multiplyExact so an overflow fails fast instead of wrapping to a bogus size), and swapped the naked new ZstdByteSize(16 * 1024)-style literals across README/docs/tests/smoke for it. Pushed as b4773df.