v0.0.28

Added

• (decode) kernel_* cargo features for per-tier CPU kernel selection (#307)

Performance

• (encode) borrow one-shot input as Fast match window (#318)
• (encode) raw-pointer reads in Fast kernel backward extension (#321)
• (decode) route RLE-mode sequence tables through the fused path (#320)
• (encode) drop redundant per-match offset coding in Fast matcher (#319)
• (encode) dedup window<->history live-region storage (#317)
• (encode) map level 19 to btultra2 + align bench labels to clevels.h (#315)
• (encode) skip decoder-table build when loading a dict for encoding (#314)
• (decode) route bulk non-overlapping copies through memcpy (ERMS) (#309)
• (encode) vectorize the Row match-finder tag scan (#305)
• (encode) lower row-matcher min match to 5 (donor parity) (#310)
• (encode) align levels 13-15 to reference search budget (#302)
• (decode) unify SIMD-copy capability detection under the CpuKernel tag (#304)

Bench

• (dict) report Rust dict-compressed size + emit compress-dict ratio (#301)