[MagicalTux]

Follow-up to #169. I built the reference Lizard CLI (2.1.0, inikep/lizard) and used it as a black box — its CLI and the two format documents in doc/, not its source — first to check the lz5 decoder, then to complete it.

1. Fix: real files over 128 KiB did not decode (predates #169)

• A compressed frame block is a level byte followed by one sub-block per 128 KiB of input, sharing one window. Only the first sub-block was decoded; the rest was reported as corrupt. With the CLI's default 4 MiB frame block, nothing larger than one sub-block decoded.
• -BD frames (blocks referring back into earlier blocks) were decoded as independent and failed with InvalidDistance.

2. Feature: every level the CLI writes now decodes

• LIZv1 codewords (levels 20–29, 40–49), including 24-bit offsets and repeat-offset tokens.
• Huffman-coded sub-streams (levels 30+). These are plain Huff0 blocks, the same coding zstd uses for literals. The module docs had left this out on the grounds that the reference picks between two decode-table strategies without recording the choice; that choice does not affect the decoded bytes, so the existing single-symbol decoder is enough.
• To share that decoder without making lz5 depend on zstd, the bit reader, FSE and Huff0 decoders move to a crate-private src/huff0/ (first commit, no code change).
• Dependent frames keep 16 MiB of history (the reach of a 24-bit offset).

3. The content checksum added in #169 is confirmed

Files the CLI writes with its default checksum decode and verify, and flipping a trailer bit fails with ChecksumMismatch. That was the one unconfirmed item from #169.

Verification

• Against the CLI: 2,640 files — every level 10–49 × -B1/-B3/-B7/default × with and without -BD × eleven inputs from empty to 8 MB — plus 1,008 more on skewed, run-heavy and 26 MB far-repeat inputs. All decode to the original, checksum verified.
• Other direction: this crate's (store-only) encoder → lizard -d passes.
• Fixtures: five CLI-written files (13–42 KiB), each decoded under four chunkings, plus truncation and single-byte-corruption sweeps over them.
• Fuzz: decoder_lz5 for 2.7M executions seeded with real Lizard streams at eight levels.
• Gates: fmt; clippy for all, all-features, lz5-only, zstd-only, lz4-only, Windows and macOS targets; full test suite; lz5-only tests; no_std builds; MSRV; docs on stable and nightly; i686 check.

Notes

• Where lizard_Block_format.md is ambiguous the CLI's output decided: which flag value means "repeat the last offset" (the document states both), and that at levels 20–29 length extensions sit inline in the literals stream rather than in the lengths stream.
• None of the CLI's output in these runs used a Huffman-coded lengths or offset stream, or a non-empty lengths stream. Those paths are implemented from the document and covered only by hand-built unit tests.
• Still Unsupported: block checksums and external dictionaries. The encoder is unchanged (store-only).

🤖 Generated with Claude Code