[switch-nz]

Part of #47 and #50: first of three PRs agreed for #47 (LZFSE, then DMG, then AFF4). #47 and #50 stay open.

What changes for examiners

• Files that macOS compresses with LZFSE (decmpfs types 11 and 12) on HFS+ and APFS are now decompressed like zlib and LZVN ones, instead of being listed with a note and no content.
• Still listed without content: files stored raw (types 9 and 10).
• engine/lzfse.py is a pure-Python decoder for raw, LZVN, v1 and v2 LZFSE blocks. It is also what DMG's ULFO chunks will use in the next PR.

Verification

• Decoder checked against Apple's reference encoder (liblzfse) on empty, tiny, text, zeros, random and large inputs; small encoder-written streams (v2, LZVN and raw blocks, and all three in one stream) are embedded in tests/lzfse_vectors.py so CI needs no dependency. The reference decoder source was used to write the layouts and FSE tables.
• v1 (bvx1) blocks, which this encoder no longer writes, are tested by rewriting a v2 header in v1 layout. That is a self-consistency check, not an independent one.
• Hardening: output cap, truncation always refused, damage raises only LzfseError; fuzz target added (3000 cases, no findings).
• HFS+ and APFS volume tests now cover LZFSE in the attribute and in the resource fork; raw (type 9) remains the "listed, not faked" case.
• Tree cache version 3 → 4 so a cached 0.8.0 listing shows these files' sizes.
• Full suite 1294 tests OK (14 skipped), smoke passes.
• Synthetic volumes only; no real macOS volume with LZFSE-compressed files was available.

🤖 Generated with Claude Code

https://claude.ai/code/session_017DPjyZZHvaivtvJ3N4hZhz

---

Generated by Claude Code