[MagicalTux]

Stacked on #145. This is kept out of that PR because of a provenance question, which is your call.

Problem

Our -lh2- encoder and decoder only agreed with each other. The real lha tool (LHa for UNIX) fails its CRC check on any non-empty lh2 data we encode. Our implementation got three parts of the format wrong:

• Length alphabet: the real format has 286 symbols, with an escape symbol followed by 8 raw bits for long matches. Ours had 510 symbols.
• Positions: the real format codes the top bits (pos >> 6) with a tree that grows as output is produced, then sends the low 6 bits raw. Ours coded a fixed 14-class bit length.
• Tree updates: LHA uses its own adaptive-Huffman update and rebuild rules, which decide every code bit. Ours used Okumura's LZHUF rules.

Fix

The encoder and decoder are rewritten to the real format. lha t now reports CRC OK on our archives, and lha p gives back the original input. tests/lha.rs now wraps each method's encoder output (lh1, lh2, lh4–lh7) in a .lzh archive and has lha verify and extract it; the check is skipped when lha isn't installed. A 208-byte lh2 fixture covers machines without lha.

⚠️ Provenance: please decide before merging

The other LHA methods in this crate are documented as clean-room. To get every code bit right, this rewrite was written from a reading of the LHa for UNIX sources (dhuf.c, slide.c), reproducing their tree-update and rebuild rules in new Rust code. LHa for UNIX has its own license with conditions, and the crate is MIT. The module docs now say the rules were "matched to the behaviour of the LHa for UNIX extractor".

Options:

• Accept it as is. The code is independent Rust that reproduces a format's behaviour.
• Clean-room it. Have someone else write a spec from that source, and reimplement from the spec alone.
• Drop it. Close this PR and make -lh2- return Error::Unsupported instead of producing data that lha rejects.

Also: no tool available here can create lh2 data, so the decoder hasn't been tested on lh2 written by another encoder. It runs the same tree code that lha accepts, so it is expected to work.

🤖 Generated with Claude Code

[MagicalTux]

Superseded by #152, a clean-room reimplementation: one agent wrote a code-free spec from the LHa for UNIX source, and a separate agent with a fresh context implemented #152 from that spec and black-box oracle testing alone. Note: this PR was merged into fix/pre-0.7.4 after #145 had already taken that branch into master, so this code never reached master.