Adaptive learned key geometry (taps + gestures)

[SHAWNERZZ]

Status: in progress — branch feat/adaptive-key-geometry.

Make the keyboard adapt to how the user actually types — for both tapping and gesture/glide input — by learning per-user key geometry and applying a light context prior. Opt-in.

Full design: docs/ADAPTIVE_TYPING.md.

Summary

• One per-user model keyed by (key, layout, orientation): mean landing offset + variance + count. Content-free (no text), incognito-gated, opt-in.
• Gestures (hard requirement): adjust the per-key sweet spots in ProximityInfo (Java-only, no native rebuild) so swipes match where your hand goes.
• Taps: bias KeyDetector.detectHitKey toward the learned center, plus a completion-derived next-character prior — all hard-capped so a clear key press can't flip to a neighbor.
• Persistence: a new table in leantype.db → rides the existing Advanced-tab backup/restore for free, backward-compatibly. Deletable via a "reset" action.
• Stats page: visualize the learned keyboard + per-key accuracy/consistency.

Build phases

1. Foundation: opt-in pref + DB table + DAO + TouchModelManager (EMA + capped geometry API).
2. Learning + gesture/tap injection (testable on-device).
3. Context prior (Layer A) for taps.
4. Stats page + reset UI + strength tuning.

Safety / privacy

Hard caps + confidence gating (no wrong literals); content-free storage; respects incognito and the opt-in toggle; validate on-device that gesture recognition improves, not destabilizes.