feat: implement Better compression level (zstd level 7, lazy2 strategy)

[polaz]

Summary

CompressionLevel::Better currently triggers unimplemented!(). Level 7 uses lazy2 strategy — significantly better compression ratio than dfast at moderate speed cost.

Analysis

C reference implementation (zstd level 7)

• Strategy: ZSTD_lazy2 (lazy matching with 2-step look-ahead)
• Parameters: windowLog=23, chainLog=23, hashLog=22, searchLog=3, minMatch=3, targetLength=48
• Key files: lib/compress/zstd_lazy.c

Algorithm

1. Hash chain table — hash → linked list of positions with same hash
2. Lazy matching — at each position:
   • Find best match at position N
   • Find best match at position N+1 (with 1 extra literal)
   • Find best match at position N+2 (with 2 extra literals)
   • Pick the best overall (longer match justifies extra literals)
3. Common-length caching — when traversing hash chains, skip bytes already compared (critical optimization from C code)
4. Row matching — ZSTD_RowMatchState with SIMD tag matching for faster chain traversal

What needs to be implemented

1. Hash chain data structure — position chains indexed by hash
2. Lazy2 match finder — 2-step look-ahead with price comparison
3. Match length comparison — efficient byte-level comparison with early exit
4. Chain depth limiting — searchLog=3 means max 8 chain entries checked
5. Rep code integration — check repeat offsets at each lazy step

Acceptance criteria

• CompressionLevel::Better no longer panics
• Roundtrip + interop tests pass
• Compression ratio within 10% of C zstd level 7 on silesia corpus
• Speed slower than Default but ratio measurably better

Dependencies

• feat: implement Default compression level (zstd level 3, dfast strategy) #5 (Default level) — shared infrastructure for hash tables and match counting

Time estimate

4d

Blocked by

• feat: implement Default compression level (zstd level 3, dfast strategy) #5 (Default level — shared hash table and match counting infrastructure)