[enaboapps]

Closes #1033

Stacked on #1030 (base claude/1029-learn-from-typing). Retarget it to main once #1030 merges, and don't delete that branch before then.

This uses the published switchify-prediction v0.2.2, so the AI suggestion words (slots 4–6) arrive 3–5× faster. For the same context they are the same words. The worker protocol (2), model and tokenizer are unchanged.

Changes

• scripts/prediction-neural.json:
  • The worker archive URLs, archive hashes and per-file hashes for all four targets are regenerated from the v0.2.2 release. Each archive was checked against its published .sha256.
  • revision (the converter and worker provenance) and library_revision are now b873c88 (v0.2.2).
  • v0.2.2 archives also contain the companion CLI. It is deliberately left unpinned, and unpack only extracts pinned files, so it is never extracted or shipped. The file count per target is the same as v0.2.1.
• src-tauri/Cargo.toml: both libraries are pinned to b873c88, which already includes the learning changes from Learn from typing to personalise word prediction #1030.
• scripts/check-packaged-prediction.mjs: expects worker BUILD.json version 0.2.2.
• Docs: neural-prediction.md and six-predictions.md describe v0.2.2.

Measurements (from switchify-prediction #26)

Worker	Median before → after
AVX2 (Windows test machine)	581 → 121 ms
Portable (Windows test machine)	1,257 → 424 ms
macOS ARM (CI, portable)	690 → 151 ms

Validation

• npm run lint and npm test (295) pass. cargo fmt --check, cargo clippy --all-targets -D warnings and cargo test (758) pass.
• The converter source (quantize.rs) and model-bundle.json are byte-identical between v0.2.1 and v0.2.2, so the pinned model hash still holds.
• The archive inside BUILD.json reports commit b873c88, version 0.2.2 and protocol 2.
• Not run locally: npm run prediction-model. On this Windows development machine, real-time scanning blocks the staging-directory rename for freshly written worker executables (EPERM), so it can't complete here. CI runs the full fetch, verification, conversion and packaged checks on Windows and macOS runners.
• Manual check: with signed builds, confirm that AI suggestions appear quickly and that Retry predictions still works.

🤖 Generated with Claude Code