[NotOfficer]

Adds a net11.0 target. On .NET 11 the built-in decompressors use the new span-based decoders (DeflateDecoder, GZipDecoder, ZLibDecoder), and CompressionAlgorithm.Zstd is registered with ZstandardDecoder. net10.0 keeps the stream-based implementation.

Behavior

Every built-in decompressor now follows the rules of BrotliDecoder.TryDecompress, on both target frameworks:

Input	Before (Deflate/Gzip/Zlib)	Now
destination too small	true, truncated output	false
incomplete source	true, partial output	false
bytes after the compressed data	ignored	ignored
several gzip members / Zstandard frames	gzip: all decoded	all decoded

TryDecompress reports bytesWritten = 0 whenever it returns false. Before, a failing handler's partial count leaked through, e.g. Brotli with a too-small destination.

Breaking change: callers that relied on truncated output for a too-small destination or an incomplete source now get false, and Decompress throws InvalidDataException.

Decisions worth a look

• Instance API instead of the static TryDecompress methods on .NET 11. The static methods also fail when any byte follows the compressed data. The instance API stops at the end of the data, which keeps padding and trailing bytes working as before.
• Gzip members and Zstandard frames are decoded in a loop with Reset():
  • Gzip uses the GZipStream rule from .NET 10 (Inflater.ResetStreamForLeftoverInput): another member starts if the rest begins with 1F, followed by 8B when a second byte exists.
  • Zstd follows ZstandardStream: another frame starts if the next 4 bytes are a frame or skippable-frame magic number.
• A full destination gets a one-byte scratch buffer. The decoders report DestinationTooSmall for an empty buffer even when nothing is left to write, e.g. an empty trailing gzip member or a skippable frame.
• net10.0 detection (TryDecompressStream and SourceStream):
  • DeflateStream only reads from its base stream while the inflater is unfinished and out of input. So SourceStream, a small UnmanagedMemoryStream subclass, flags any read past the end as an incomplete source.
  • A one-byte read after the destination is full detects output that didn't fit.
  • GZipStream reads once more after a member that ends exactly at the end of the source. A sentinel byte (0xFF, which can't start a member) answers that read. Deflate and Zlib don't get one.

Other changes

• Test packages updated (Microsoft.NET.Test.Sdk 18.11.1, xunit.v3 4.0.3, xunit.runner.visualstudio 4.0.1, coverlet.collector 10.1.0), and src/global.json runs dotnet test on Microsoft.Testing.Platform.
• CI installs .NET 11 and now also runs build and tests for pull requests to master. Pack, artifact upload and the NuGet push only run for pushes to master.
• .gitattributes with * text=auto. The repository already stores LF, so no file changes.
• JsonUtils no longer has an unneeded unsafe modifier.
• The package version stays 3.2.0, so the push to master after merging won't publish a new package (--skip-duplicate).

Validation

New tests in IDecompressorTests.cs (the 3.2.0 tests are unchanged):

• round trip with 256 KB for every algorithm
• trailing bytes are ignored
• failures: invalid data, empty source, incomplete source, too-small destination
• gzip members, Zstandard frames (net11.0)
• the .NET 11 span encoders (net11.0)

Checks run locally on Windows with SDK 11.0.100:

• Tests: dotnet test -c Release on this branch in a clean worktree: 90 passed on net10.0 and net11.0, 0 build warnings.

• Same results on both target frameworks: 61,120 generated inputs, decoded on .NET 10 and .NET 11: every prefix of the compressed data, padding variants, destination sizes, concatenated streams, flipped bytes. 0 mismatches. The net10.0 build running on the .NET 11 runtime (roll-forward) gave 0 mismatches as well.

• Mutation check: 13 deliberate bugs in IDecompressor.cs, run against the new tests:

  • Examples: dropping the probe read, the sentinel, the gzip loop or the scratch buffer.
  • 12 make tests fail. The 13th, sentinel 0x00 instead of 0xFF, behaves identically: any value except 0x1F works.

• net10.0 overhead vs. 3.2.0 (Stopwatch, exact-size destination):

  Payload	zlib	gzip
  60 B	0.69 → 0.74 µs	0.77 → 0.87 µs
  64 KB	5.65 → 5.58 µs	7.93 → 8.29 µs
  4 MB	250 → 244 µs	395 → 401 µs

🤖 Generated by Opus 5.5 (xhigh) with Claude Code in T3 Code