[enaboapps]

Closes #1029

Word prediction now learns from what the person types with the scanning keyboard, so their own words, names and phrases are suggested sooner.

What changes

• Setting. Adds Learn from my typing under Scanning › Keyboard › Suggestions, on by default.

  • It is stored in AppSettings beside autoApproveAccountDevices. It is per computer, excluded from synced settings, and the synced schema is unchanged.
  • Forget learned words deletes what was learned, after a confirmation step.

• Where it is stored. The isolated prediction worker opens prediction/personal.sqlite in this computer's app config directory and passes it to switchify-prediction.

  • If the file can't be opened, prediction keeps working without learning.

• What is learned. Only the existing prediction buffer is learned from, which is successful Switchify keyboard input:

  • A sentence is learned once, when the next one begins, so a fix just after the punctuation is learned corrected. Backspacing into learned text relearns that word whole.
  • Enter/Tab learns everything typed so far. The keyboard now queues a private Edit::End, which keeps the last word, instead of a Reset.
  • Other context loss learns only the completed words, not the word in progress.
  • Text about to be clipped from the 512-character buffer has its completed words learned first.

• Changing the setting bumps a learning generation, so the running worker and any spare are replaced at the next keyboard update.

• Forget sets a pending flag, replaces any worker holding the file, and deletes the file plus its -journal, -wal and -shm. It finishes as soon as no worker holds them: immediately, on stop(), or before the next worker starts. No worker reopens the file while the delete is pending.

• Library pin. switchify-prediction is pinned to Make learned words count, boundedly, without a second baseline copy switchify-prediction#30 (stacked on feat: redesign desktop setup UI #28):

  • Incremental learning: about 8 ms per sentence instead of about 190 ms.
  • Bounded weighting: one learned use counts as up to 500 built-in uses, less as more is learned.
  • Typo protection: a new or rare word counts once learned twice, so a single typo is never suggested.
  • Results on a synthetic person: learned words reach the suggestions for 46% of queries instead of 1%, and no misspelling was suggested.
  • Memory: the library no longer copies the English model, so the prediction worker's peak memory falls from about 362 MiB to 134 MiB.

  The worker release assets (v0.2.1) are unchanged, and library_revision was updated to match.

• AI slots. The AI slots don't learn themselves. They exclude slots 1–3, so they still offer different words.

• Docs. word-prediction.md gains a Learning from typing section. The neural, six-slot and keyboard docs are updated, and so is the settings note, which said typing "is not used for learning".

Typed text is never logged or synced. The file holds counts, not a transcript, but it is unencrypted, and the settings note says it can include names or passwords.

Validation

All checks from AGENTS.md pass: npm run prediction-model, lint, test, build, cargo fmt --check, cargo clippy --all-targets -D warnings, and cargo test (754 + 7 tests).

New tests, all using fake adapters with no desktop input:

• Worker: a sentence is learned once when the next begins; a correction after the punctuation is learned corrected; End keeps the last word; lost context skips the word in progress; clipped text is learned in whole words.
• Model: a real SQLite personal file ranks learned words, persists to the next worker, leaves the baseline byte-identical, and learns nothing with learning off.
• Main process: Enter/Tab queue End; a spare started for another personal file isn't adopted; closing passes the personal file to the spare; Forget deletes the file and its journals, only when asked.
• Settings: learnWordPrediction defaults to true for stored settings, is excluded from the synced document, and survives an incoming sync.
• UI: the toggle saves app settings, not the scan config, and Forget calls the backend only after confirmation.

The new tests originally started enough PowerShell fake workers to starve cancellation_kills_and_reaps_a_blocked_worker after a rebuild. They now start only what they need, and the suite passes on repeated runs straight after a rebuild. One unrelated account-sync UI test flaked once during a concurrent build and passed on three reruns.

Manual check still to do: with signed builds (npm run macos:run on macOS), type sentences and check that learned words rank higher after reopening the keyboard. Then use Forget and check the file is gone.

Note: #1026 (docs) edits some of the same doc lines. Whichever merges second needs a small rebase.

🤖 Generated with Claude Code

[enaboapps]

Pins switchify-prediction e5223aa (#30). #30's final head b1904a9 differs only in docs, so the code is identical. Re-pin to the merged main commit if #28 and #30 are squash-merged.