[MagicalTux]

Replaces #146, which reproduced LHa for UNIX's tree rules from a reading of its source (a licensing problem for this MIT crate).

Problem

Our -lh2- encoder and decoder only agreed with each other. The codec used a 510-symbol length alphabet, a 14-symbol bit-length position code and Okumura's LZHUF tree rules, none of which is the -lh2- format. It could not decode real -lh2- data, and LHa could not decode what it wrote.

Clean-room process

• Spec side: one agent read the LHa for UNIX source and wrote a functional spec of the format: prose, math, tables and test vectors, with no code, pseudocode or identifiers. It checked the spec by encoding streams with its own throwaway model and having LHa for UNIX decode them. It also built LHa for UNIX into a standalone binary to serve as a black-box oracle.
• Review: the coordinator read the spec for leaked code before passing it on.
• Implementation side: a separate agent with a fresh context wrote this change. It had only the spec, the oracle binary run as a black box, and master. It never saw fix(lha): make -lh2- match the LHarc format (needs a licensing decision) #146, the LHa source or the spec writer's context.

Fix

• Char tree: 286 symbols covering literals, lengths 3..=31, and an escape followed by 8 raw bits for longer lengths.
• Position tree: codes P >> 6 and grows by one symbol per 64 output bytes. It keeps LHa's root sentinel and its separate update counter, and the low 6 bits follow raw.
• Window: 8 KiB, filled with spaces at the start.
• Limits: the encoder caps lengths at 256 and distances at the bytes already output. The decoder accepts lengths up to 287.

Tests

• Spec vectors: initial codes and an adaptation trace; five streams encoded and decoded bit for bit; a 4 MB random token stream that rebuilds both trees, checked by SHA-256.
• Fixture: a new oracle-verified fixture, checked against the decoder only.
• LHa for UNIX conformance: opt-in. LHa for UNIX must test CRC-clean and extract our lh1, lh2 and lh4–lh7 output. The binary comes from COMPCOL_LHA, or from lha on PATH only if it reports itself as LHa for UNIX. Otherwise the test skips; Homebrew's lha is Lhasa, which can't decode -lh2-, so it skips there.
• Fuzzing: the decoder_lha fuzz target now covers lh2.

Results:

• With LHa for UNIX 1.14i set through COMPCOL_LHA: all pass.
• Without it: all pass, with the oracle test skipped.
• cargo fmt --check and clippy with -D warnings: clean.
• Cross-check: fix(lha): make -lh2- match the LHarc format (needs a licensing decision) #146's LHa-verified fixture decodes correctly with this independent decoder.

Not verified

• No fuzzing was run; the fuzz target only builds.
• The MSRV build wasn't checked.
• There are no historical -lh2- archives, and LHa for UNIX can decode -lh2- but not create it. Compatibility is shown against its decoder only.

🤖 Generated with Claude Code