Use switchify-prediction v0.2.2 for faster word generation #1033

switchify-prediction v0.2.2 is published. Its workers generate the AI suggestion words 3–5× faster, with identical words for the same context: a median of 581 to 121 ms with AVX2, 1,257 to 424 ms portable, and 690 to 151 ms on macOS ARM CI. Generation also now sees the sentence as typed. The worker protocol, model and tokenizer are unchanged.

Pin the v0.2.2 worker archives and per-file hashes in scripts/prediction-neural.json. Move both Rust libraries to the v0.2.2 commit, which also includes the learning changes already pinned by #1030. Update check-packaged-prediction.mjs to expect version 0.2.2, and update the docs.