Two lossless binary-size reductions. The musl inspect example goes 778,528 → 770,944 B (−7,584, −0.97%); macOS is unchanged (segment quantization), so the README's "621K" stays accurate. All tests and the JS-equivalence proptests pass.

Zopfli for the codegen blobs (−6,526 B across all 10 blobs)

Compression happens only in the xtask codegen, and DEFLATE decoding doesn't care how a valid stream was produced. So the encoder is switched from flate2's Compression::best() to Zopfli — a deflate-compatible encoder that uses iterative optimal (shortest-path) LZ77 parsing to squeeze a few percent more from the same format.

• The output is a standard raw-deflate stream, so the runtime flate2::read::DeflateDecoder reads it unchanged — no format or decoder change.
• xtask is publish = false and is not in the crate's runtime dependency tree (cargo tree -e normal shows no zopfli), so no consumer-facing dependency changes.
• The extra build-time cost lands only in offline codegen. Default 15 iterations is the sweet spot (100 iterations saved only ~60 B more for ~50× the time).
• Biggest blobs: pair_indices 44,567 → 42,556, percentages 39,019 → 35,623, feature_matching 21,188 → 20,578.

Region percentages: 3-byte byte-plane + drop PERCENT_RANGES (−1,056 B)

The region percentage deltas now use three fixed-width byte planes (low/mid/high) — the same stream-split scheme already used for the pair indices — instead of postcard varint.

• Being fixed-width and element-addressable, the percentages reuse PAIR_RANGES (pairs and percentages share identical per-region counts), so the separate PERCENT_RANGES offset table is removed (−964 B of .rodata).
• Also drops the decode::<u32> postcard monomorphization. RegionData collapses from 4 fields to 2.

🤖 Generated with Claude Code