2026-09-14 報 release. 14 September 2026

The hybrid typo-retrieval layer is live across the ecosystem

A 0.481 MB character bi-encoder now retrieves the twenty nearest vocabulary entries for a misspelling and the fastText full tier rescores them, with the layer wired end to end: registry v1.7.0 carries the artifacts, kotoshu 1.0.3 fetches them with one setup flag, and the Rust crate and WebAssembly package expose the same engine.

The registry v1.7.0 release promotes the typo bi-encoder from an opt-in mirror to a first-class artifact pair, generated by the descriptor-driven release flow with the validator enforcing that the model and its character vocabulary travel together. kotoshu 1.0.3 adds kotoshu setup LANG --typo: one flag fetches the pair plus the full tier, reports the layer as cached on later runs, and arms it behind KOTOSHU_TYPO_RETRIEVAL with a byte-identical passthrough when off.
The engine reproduces the frozen benchmark exactly through the shipped path - top1 0.118772 and top5 0.269829 on the 2,509 real English pairs, identical to the measured verdict - at 1.8 ms per query after the index builds eagerly with the Ruby GVL released (a 25-second derivation no longer pauses any thread). The pure-Rust forward pass matches onnxruntime to cosine 1.0. kotoshu-native 0.1.1 on PyPI now carries all 21 platform wheels, and the crate 0.2.0 and @kotoshu/wasm 1.1.0 cut the same surface for Cargo and npm. Prebuilt 26 MB KTM1 matrices - one per full-feature language, served by the registry - arm the same engine in about half a second instead of a 25-to-45-second index derivation.