Every compressed blob was loaded with ad-hoc OnceLock + inline postcard::from_bytes(&decompress_deflate(include_bytes!(…))) boilerplate — repeated and slightly varied across ~10 call sites plus the two generated loaders. This consolidates them behind three small helpers in compression.rs:

• LazyData<T> — a blob decompressed + postcard-deserialized into T on first use and cached. For values consumed as-is: the feature/region version tables, the per-browser version order, and the generated key tables.
• LazyBlob — a blob decompressed to raw bytes and cached, for data decoded by hand at the use site (the byte-plane and run-length blobs, where a generic deserializer would bloat the binary).
• load<T> — one-shot decompress + deserialize, shared by LazyData and by the few sites that transform the decoded value further (node versions, the browser map, the region pair table).

Call sites lose their hand-written accessor functions and become uniform static X: LazyData<…> = LazyData::new(include_bytes!(…)) / X.get(). The remaining OnceLocks now cache only derived/transformed values (e.g. the browser FxHashMap, version aliases), and their blob loads route through load().

No data/behavior change, small size win

Blobs are byte-identical (regenerated). Sharing the loaders de-duplicates the per-site monomorphized closures, so the musl inspect example drops 764,160 → 761,920 B (−2,240). All tests and the JS-equivalence proptests pass.

🤖 Generated with Claude Code